ISTANFORD ARTIFICIAL INTELLIGENCE PROJECTMEMO AIM-161

_i -_- _ _ _ ' e ISTAN-CS-264-72

ANARTIFICIAL INTELLIGENCE APPROACHTOMACHINETRANSLATION

BY

YOR ICK WILKS

--_ SUPPORTED BY

OFFICEOFNAVALRESEARCH

AND

ADVANCED RESEARCH PROJECTS AGENCY

ARPAORDER NO. 457

a

FEBRUARY1972

COMPUTER SCIENCE DEPARTMENT

School of Humanitiesand Sciences

STANFORD UNIVERSITY

c

i-

i.

Lii

L

STAWORD ARTIFICI AL INTELLIGENCE PROJECTMEMO AIM1161

COMPUTER SCIENCE 9EPARTMENTREPORT CS-264

FEEWARY 1972

. .

AN ARTIFICIAL INTELLIGENCE APPROACH TO MACHINE TRANSLATION

by

Yotick WI l.ks

ABSTRACT t The paper descr!bes a system ot wnantlc analysfs andgeneratlonr Programmedparagraph length Inout In

In LISP 13 and deslmed to Dass from

reoreSentatlon,A wfdeEnglfsh t o F r e n c h da an fn&lingual

class of English InPUt forms will be covered,but the vocabulary wtll lnltlally be restrIcted to one of of a fewhundred words, Mth this subset workfng‘year (71-721t

and during ihe ourrsncIt 1s also hoped to map the IntirIlngual repreaontatfon

onto some Predfcaje calculus notationanskerlng of V0rY sfmPl0 questIons

SO as to make p o s s i b l e &eabout

The soectflcattonthe translated matter,

of the translation system kself 1s oomblete, andIts main points of interest that dlstlngulsh It from other systemsare:

(Continued on page 11)

qresented at the Conference on Natural Languagethe Courant Institute

PrOC0Sslng held aiINeW York Unlvorsfty on December 20-21, 1971,

The vlebs and conclusfons aontalned In thfs document are those of ihoauehor and should not be represented as neoessarlly reprasentjng iheofflpial pollc~c~s, either expressed or implied tof tha AdvancedResearch ProJacts Agency or the U&Government,

Thfs research was supported In part by an ONR Contract (f'J0001447=A-00491, and In Dart bY the Advanced ResearchAgencYdepartment 09 Defense (SD 183), U,S,A,

PToJeats

Reproduced In the U&A, Available from the ClearInghouse for Fede;alScientlfk and TechnIcal Informatfon, SpringfIeld c Virgfnfa, 22151,

;)It translates phrase by phrase----with faciiities for reorderingphrases end establishin essential semanttc oonnectivftles betweontherr - - - b y mapphg complex semantic StrUCtUres Of “messa 81’ ontoeach phrase, These constitute th%.lnterllngual ?vwresentat on to betranslated, This matchiog is done without the @xDiiCli: yse of 8conventional syntax anaiYsis, by taking as the apmoPriat@ vatchedstructure the "most dense 0 of the alternative struotures derfved,Thismeihod has been found highly s~ccessf~i in eariler versfkns bf thhanalysis system,

Ii) The Frenoh output strings are generated without the exdl!c!t useof a generative grammar,fhat fs done by means of STEREO'fYPES: strinOsof French word% and functtons evaluating to FCsnch words, whioh yeattached to English word senses in the dfctiona;Y and built into theInterlingual representation by the analysis routfnes,Tha gsneratfonprogram thus recefves a n interlingual representation thai alreadycontains both Frenoh output and implicit Drocedures f0, aasembllngthe output, stnce. ihe stereotypes are in effect reoursb dioceduiesspecifying the content and production of the output Word s~r~ngs,Thu?the generation program at no time consults a word diat\onarY OrInventOrY at grammar ruiW3,

It Is claimed that ihe system of notation and translation descr 1 bedii a convenient one for expressing and handling the Items of semantlojnformation that are ESSENTIAL to any effective MT system,1 dIsoussIn gome detail the semantio informatton needed to ensut8 the cortwichoice of output preoosIt\ons in Frenchja vital lratter lnadeauateiytreated by virtually all previous formalisms and proJects,

1,0JIntroduotion

1 call khat follows an Artificial Intelifgence (AI) approach to iheproblem of i”lachine Translation (MT) for five reasons!

i)when fully developed the syd'tem to be described for representIngnatural language WI II contain withln itself two meihods f o rexppessingIinguistlor

the content of any given utterancetone logical, the otherin a broad sense of that term,It is at the Dreseni time

an outstanding quesjton within Artificial Intellrgence Which of thesegeneral approaches fs the most suitable,In that the Present sYs;emhas both representation capabllitles, it should be able to GowaretheIT with a view to throwing some light on this importani dfspute,

211 have argued elsewhere Cl41 at some Moth that the spaoe ofmeaningful expressions of a natural language cannot be determined ordeciaed by any set of rules whatever------In the way that arlrnoat al IIinWJlstic theories lmDlicitlY assume CAN be done,That 1s because, inoomnlon sense terms, a sp8ak9r a/ways has the optlonstritW Of

to .MAI(E a n ywords meanlnsful b y the use of

definitions, iiowever, any working system ofexPlanatfons and

linguistloimplicitly swcMy a class of acceptable exPressions

&tles does

indirectly, a class of unaoceptable ones, The only way o Ia n d SO,

oombfn'fnethese two facts of life Is to have a modifiable system of lfnguisiiorules, which was Implemented in anVersion of the Present system Ci33,

elementary way tn an earlie;

3)Another aspect of the AI armroachr if one can use that dhrass, hasbeen an attraction to methods consistent with what humans THINK thet;methods of’ procedure aret as distinct from more formally moiivacedmethodsWnoe the a&action of heuristics in, say8 AI approaches totheorem ptoving,The pvc3ent system fs entirely semantics based, fnihat it avoids the explicit use of a oonventtonalat both the analysis and the generation stages,

itnsddc sYn:ax

tQPUt IIn the analysis of

syntax is avoided bY a template systemrthe use of a set ofsgantlc forms that seek to Dick UP the message conveyed by ihe inputstring, on the assUmDtfon that there is a fairly well defined set ofbas i c messages that people always want to convey whenever jhey wrjteand speak; and that ln order to analyse and express the content ofdiscourse it is these sfmple mesSages--suCh 89 that 'a og;tajn thinghas a: oertain part' for example-- that we need to locate,0vera-l I

Again,therepresentation of c0mPl.e~ sentences Is that of a lInea;

seaUenC@ Of these message forms fn a real time order ,inte;&ated byc0ncePtUal tb3, rather than the hierarchIcal treeareferred

struCturebY IlnWsts, From the very c o m m o n s e n s e f o r m s o f

excvesslon I have had to use to express this method of attack ft wfl ibe seen that the method itself is one olose to ordfnary intuitionsabout how we understand, and somewhat distant from the concerns offOrK8l gramarlans,

2

r))The French generatIon Is done without the 8XPlloli uue of agenerative Qrammar, In the conventtonal senae,The tnterllnoual~eOresentatl0n Passed 'r0P-l the analysis roudnes to the generatIonones already contains, as Part of the coding of the Cngllsh ln~utworcls, French Bter@OtYPes ---strl.ngS of Frsnoh words and funotionsthat evaluate to French words,These are evaluatedrecUrSJVelY to produce French output, ste;eotYPas thusoonstltute both French output and PrOcSdures for assemblTng thatoutput oroperlY,No other inventory of Frenoh words or grammar rules1s ever searohed, and the stereotypes oonstituie a prinofDied WaY oiooping wfth Iinsuls~l~ dlVerSitY and frrrgularltYI--asinoe i?dlvidua!word3 have thejr own stereotyPes,*~wtthout reoourso t o whatBar-hl I le(Cl3 cal Is "bags of trfckaY

5)A POfnt related to (11 but Importantly d'ffferent is ?hat of i".he"level of understanding" required for MTlIi would oe;iainlY beun/ntelllgent to develop any level of understandfng more oomPlex_thanIs reOUlr8d for anY task, and It 1s hoped that bY the gethodsdescribed it may, be possfble to establish a level of unds&and!;ifor MT, somewhat Short of that reqU!red for questton

behavlors,Whll8 agree)nganswerfng

other fYOr0 fntell /gent with Mlohl8'8C63unexceptionable (0, we now have as a touchstone the realtzdon thaithe central opsrattons of the Intel JIgmoe a;e,,,,tranaaot1ons on ak.nokledge base”, It te hoped that for MT I Ingul&c, 0;lfngufstlcally exPresslbla, knowledge may sufffos,

It Is ths semantic approach that Is Intended to answer ihe aufteproPer OueStlOn qJhy Start ?I1 again at all?' The Senerall~ nowkt~veSUrveYS produced after the demise of most of th8 MT research Of Eh8Fifties In no way establIshed that a whol(Y new aPPrOaoh I /ke ;hePresent one was foredoomed to fall --on(Y that the methods tr!rd sofar had In fact done so,At this distance In time #it is easy to beunfair to the memory of that early MT work and to overexaggerate ftssimple agumPtIons about language,But the fact iemalns that almost all0-f it was done on the basis of naive syntactfc analYsts and withbUGany of the developments in semantic structuring and descc<lption thaihave been the most noteworthy features of recent linouistfe advance,

One lord of warning 1s appropriate at thts Point about ihe s8man$tcmethod and Its relation to the form of thfs Paper,ThIs 1s %endsd tobe a Practical note, conoerned to describe what 1s being )done in apartlCUlar system and researoh ProJSct, so It fs not conoeined toargue abStractlY for the v a l u e o f systems based on conoeatualconnectlonstthfs has been done elsewhere by Writers such asSlmmansCi23, Quilllan[93, KleinC3!, Schank Cl11 as well as my8elf.I arr. not concerned to argue for a general methad, nor Shall 1 set outmuch In the way of ths now fad I lar graph structUr8S &king theitens of example sentences In Order to display th8fr 'real s&/N&'for WY PUrPOSeS,I a m concerned more to display the lnfoimatlonstructure I uge, and the manipulations the sYstsm appibs to 08rtafnlinguistic examples fn order to get them into ihe prescribed farm fortranslation,The dlsDlaY of conceptual Or dependency connections

3

between items of +a( text WI I I only be made In CIRSBS whereunnecessary obscurity or oomplexltY would be introduced by d!sQlaYbWthe same connedons between items of the lnterllngual represeniatlbn,

It has beco,rle fashlonable reQentlY to olalm that ‘dlctknaky based’systems cannot find a place wlthln AL1 would like to a;gus at theoutset of tnls page; that this vie% PerV4dW though r&rely made0x01 lch la an unhelpful one8 and can only inhibit PrQgfesS on iheunderstanding of natural language ln an AI oontext,

The rise of this view can, I thtnk, be correlated with the freshlntereat being generated among llngulsts and other8 by new attsmDtssuch as Montague~sCtJ 8 to produce a formal logic cadable a;reoresentlng rather more of the forms of language than the classicatterrpts o f Russell, CarnaD, Relchenbach et a l , Thrr lmcd lcttarsurent 908s a s follows: that logical strucrture provides the realstructure of language, and there Is no Place In a lodc fQr adlctlonary, henbe,,,,,,

But In so far aa any premise of this argument Is m a d e pie&se It canthen be seen to-be h\ghly mlsleadlng, If not downrlght false,Therelation o f forma I logfc to language Is and aNays haa been a muuhdfst?Uted matter and cannot be discussed here In any d@takBut anyadequate logic must oontaln a dlctlonaty or Its equivalent if It Is.to handle anythlng more than terms with naive denotatfons such as'chair', AnY system of analysis that Is to handle sentences conkaklng8 sah 'hand' Is going to n08d to have avallable In some form suchInformatIon a s that a hand Is a part of a body, and iha It 19somethlng that only human belngs hava,It does not matter whether thfslnformatlon Is explfcitly tied ta a word name fn the form of markeis,or Is expressed as a series of true assertIons a dlQtlona;y is whatIt 1% and If the jnformatlon Is adequately WWressed ii must beposs/ble to construct efther of those forms from the other, Just a san ordinary English dfctlonary expresses informatlon in a mixture ofboth forms, On the whole, the v expllclt dfctfonaryft 1s a m o r e

e economical form of expresdOn,

Those who attack 'dictionary based' gystams do not swt! io SW thatmatters could not be OtherwIse,, Pressed for alternatlves thai expresstheir p o i n t o f view, they are now Drone tQ refer to WlnogiadC161,BUtthat is absurdlWinograd9 work oertalnly contains a dfotionarY,thefadt Is not as obvious as it mlaht b e because Qf the highlysimpllfiad Universe wfth which he deals, and the direct denotatlonaln a t u r e o f the words It contalna,But my Dolnt holds even withIn thatslmpllfled world,to see this one only has to read WlnoQ;ad'a workwith the questlon in mind:how does the system know, say, that a blookIs 'hanaleable', The answer is out quite clearly In a texi figuresbymeans of a small marker dlctlonary of course,

MichieC63 has WrItten Of “., the mandatory relationshio, ignored bysome coffputational Ilngulsts, between what Is monadfc, what isstrUctural~ and what It eplstemlc,ft ln connexlon with hls claim that

Wlnograd's work oonstltutes "the first aucoessCul s0lUttbn of ;hsmaohlne translatlon oroblem", But It may not be mere fpnoranoe Onihe port of myself hare, and others elsewhere, In vfew of the faoithat the dlstlnction between what Is q90!stamlc'f and wh&t la noi----I thlnk Mlchie means by that word noonoerned wlth the real wo;idiather than with I anguage? a rather soeo'fal and non&~adltlonalmeaning -“-“ml ; g by no means as olear as,he thtnka# It 8eems to methat the onus of PrOOf IS on the beilevers - - t h a t knbwledea abi+gthe real world IN SOME: STRONG SENSE OF THOSE WOROS lu naosFBssapy forIlngulstlc tasks IIke MT,It 1s usual to refer J as Mfqhle does, toexamDIes Ilk@ Whos;ad4s dlstinotlon between the anaohoraar tn "Thecity Council refused the women a Permft because they feared violence"and "The Cfty Counofl refused the women a aerm'ft bsoause iheY Wereoommunlsts*~,But if the eDlstemlo bellavers mean &Y "knowledge of iheworld" the 9flnductlve knowledge of the average marP then ihr)y a r ebefng over parochjal In aooepting such axamP/es at facce v lue;li alldeoends on whether ihe Clty Counall 1s WashIngCon's or Pak ng'srt andan intelligent system might be perfeod~ rleht to r@fUss io assignthe an(aphora In !&uOh trlok e x a m p l e s a t a l l ,

1 am not suggesting, though, that the manloulailons to b e desorlbadhere are mere IY ‘dktfonary based% If that Ia to be &ken ia maranhaving n o theoretfcal presuoposltfOns,There are in $a& th;ee!-mo;:;;;t llngulstfc DrWWDDOdtbnfJ On which the foliOWh?g analYsta

: namely the use of templat4s for analysis, andfo9r generation referred to above and described In data!1

tereotYpes

of the PaDer, &idt n the body

In addItIOn the orfncibls c to be develooed below,hat by bulldlng up the densest, or most connsotedr re ;essntntlonthat lt can for a ~!eoe of language the system of anaM 8P will b esettb the word senses and muoh of the grammar right,Wha< f mean by#density of oonnectTon )) here will be the subjeot of muoh thatfollows,

1,l)SDme other DrdfmlnarY cwastlonsm

The last seotlon was concerned Wth the cruestlon of the Oontmt ofihe I n f o r m a t i o n reaulred to do MT,Certaln kinds of fnfo;mrtjondictate their form Of exDression;if It is agreed by all a&{les thatto -do MT we need to know the fact, that hands have four f'fngersr thensome: form of rePrasentatlOn at least as strong a6 set theo;y or Ihepredloate oalculus wtll be needed to express that &&The need foif a c t s o f that sort fs a disputed one, but It is beyond dirduts thatwe shall need to know that a saYI a soldle r 1s a human being And anImbortant guestlon that artsas is, what form of reoresen&t o nt tsnecessary for facts of that sort,

This DrOJeOt IS intended to DrodUce a working artifaot and not tdsettle intellectual Questlons,Nsvsrtheless, because the &&OrY has cbeen gone over so heavily In the Past Years and because the ausstfonsstill at issue seem to cause the adobtlon of VW-Y definite Dofnts a*view by observers and - parttolpants alIke, ft 1s necessary toremarks on oertaln matters before any detalled MT work can

make

5

started, In particular, different VIWS are held at the prerrent dmeon the questlon of whether the intermediate b8tWsen

I k- two languages for MT should be logIcal or Iingutstjc in foim,I -

What the key words In that last sentence, **IogkaP and "l'lnguIat]c*V

%

L-

L.

k--

I

I-

i

L

L

I actually mean ls not as clear a’s might appear; for examole~ theyar* almost certainly not exClU~?Ve methods of attacking the p;oblam;fn that any "tOghal coding “of text will regulre a goad dsal of’ whatfs beat called llngujstlc anaiYsis in order to get the teit lnto ihehaulred logIcal fo&such 83 coaing '4th sense amblgufiy, Clausedependency and S O on,On the othe r hand few llngutatlcally orianiedpeople buould deny the need for some analysis of the logteal &atfonspresent in the discours;eF;aF; analysed, Howwer, for the PUrOOsesof the present DroJect asaumpttons may be made $afelYt(a)khatevet linguists and phllosoohers may say to the contrary, fthas never been shown that there are llngulstfo forms whose meaningCANNOT be reDresented ln any lostcal system whatever,eXatW0, lfngutsts o f t e n broduce kinds of

so, toitnfarence inference

Properly made but not catered forcalcullrsuth as -the "and so" Inference in

In oonventlonal” I

exlstlngfelt tired and went

home", but nothing follows to the effect that suuh an fnfe;enca oouldnot be coped with by means of a simple and approortate adjusimant fnrules of inference,

(b)Nhatsver loglclans may bdiW8 to the Contrary b It ha3 never b e e nishohn that human befngs Perform anything ltke a lostcat i+analatTonwhen they translate sentence3 from one language to anothe;, nor haglt ever been shown that lt is NECESSARY to do that In orda; tbtfans late m6Chanlcally, T o take a trlvial e x a m p l e , if one wants totranslate the Engl fsh **W, then for an adequate LOGICAL iian8let!onone wfll almost certainly want to know whether the bartlcular uge of~~1s" In question Is best rendered lnto logfc by !dentltY, setmember3hip or set tnclu9lon,Yet for the PurDOses of translattng anEnglish sentence contalnlng " 1s" Into a closely related languagesuch as French It is highly unlikely that one would ever wani io makespy such distinctron for the OurPOse immediately In hand,

The above assumptions ln no way close off discu33lon of the auastiOn3outstandlnW they merelY allow construuthspaFtl4SJIa

work to orooeed,Inr Dhikk3oPh lcal dlscusslon should be continue on WexaGlu

what : the I lnwlst is trYInS t o say when he says that there arcsIi;;;.istic forms and common senses inferences beyond the goobe of any

and 0) exactly what the loglcfan is trying to say when heholds ir! a strong form the thesis that logloat form 1s the basjs o fbrain codlngr Or ls the aOOroOrIaCe bas!s for oamputing eve; natuiallanguage,

There are also Interesting comDarisons to be made on this doint among0ontemPorarY academic develobments, andtogether at the present time of the

ln partlculab $he drrwlngInterests and aDproaches of

hitherto senarated work:the extended set logic ofexample that he claimed coped with llngufstio structure

MoniaQUa forbetter than

did MIT Ilnguistlcs, and, on the other hand, the lingukio work ofC,Lakoff C4Jwhioh olafms t h a t t h e transformatlonallsts in general a n dChomsky In particular ALWAYS WERE seeking for some quits o~nventiona!notlon of logtoal form and should have faced up to the fao$ In theirwork, But those fnterestlng questions ate not Issues he 8, beoausethe aim of the present proJect Is to produoe a smal I nit iaoi;f thalnot only translates from one natural language to another but 1s also,p0tentlallY at least, capable of some loglc tianslatfan a n d s oadmitting of cuestion answertng and the addltlonal Vndsrstandlng*~that that imblie3,

so, given a commftment to a question answering taollltY as wall as an~7 oner there can be no real problem about the ooe%!stenoe of the twoforms of coding, logical a n d Itnaulstlo~ wlthfn a single sysgembeoause al I but the most dogmatlo llngulsts would admft ihe neecj ofsome log- ioal analyfs wlthin any reasonable auestfon answeringsystem,However~ the coexistence might alao preclude what one wculd fnfantasy llke to have, namely a way of testing against each g$hfr haloglclst and lin_PuistI0 hypotheses about MT,Suoh a test would beprecluded because any logIoal translation (In the 3ense oftranslation Into logic) wlthln such a system would have muoh of $,hework done bY the (IngUfstic analysts thaii the systy alsocontained,So there could be no real oomparlson of the two oaths

ENGLISH---~---PREDICATE CALCULUS REPRESENTATIONI~-~~~FRENCHENGLISH --------LINGUISTIC CONCEpTUALItAtION-~~-~~-FR~NCH

because the first Path would also oontnln aulte a bit of ihe Irt&in order to get the natural language 1nbut Into logfoal fbim, But fiylsht, as I discuss below be possible to get translated output bytwo different Paths in a'f31ngle swtem and so give some ;e!n to {hanotion of eXperImental compari3on.

It is important to be clear at this Point that the dlspuce betweenthe loglcists and the I Ingulsts 13 often Unsymmetrical In fo;m,One holdi CIn. a strong (oglolst thesis about MT asserts, Tt seem8 tame8 that a PC reDres0ntatlOn Is necessary fo

fthe task,The (IngutsF

of correspondingly strong commitment denies th s8 but does not alwaysassert that a lfngulstk reoresentatlon is necessary, He may adm&that:a logical representation is sufflclent, denyfng only {hat lt 1snece3sarYexplicit mo;e

He might argue that a logical reoresentatlon m a k e sinformation 1n the input text than 1s necessary, BY

this he means simply that It Is harder to tianslate intb a logioalnotatlop than most 1 insuistio ones--a faot well attmbd to byiesearoh proJect3 o f the past -----In that more access tod~OtiOnaries and lo&v3 of Informatlon OutsIde the text ytsel% 13necessary' in the logical tradatlon case,

This is what I mean by saying that the logic translation may oontainmore information than a semantic oner even though the text tiansiaiiedcan Clearly contafn onlx the lnformatlon It contafns,The addkfonalfnformatlon comes from the extra-textual d~ctlonarles and axfoms,

L-

7

The loglcist on the other hand, WI most lfksly deny tha$ alin$uistic repr~sen~atlon is even sufficient for MT,

However, one must be a little cautious here about ihe admissfon thate loglca! coding contains more Information than a Ilngulstjc-semanilcone? a s those terms are usuallyrepresentation

understood,Any Ilngutsilc1s going to tie some such marker as MAN or HUMAN to a

woia llke "soIdler", so that when “soldier” ocours in a <eit thatL 3uStem IS going to be Just as capable of InferrIng that a'man is

being talked about as Is a system that contains an expllo/t predicate.-calculus axiom (Vx),SOLOIER(x)>MAN(x),

What is usually meant by an admission that a logloal representat]on- may contain more Information than a surely llngufstlo o n e cbnoeins

c the notation for v&able identffioation (as tn the Wlnogiad "women'*exarr.pie above) and the exlstentlal cruantiffer notatton,Thougnr aQafn

>- there is no reason tokannot be adapted to

think that a llnguistfo marker notatjoncope with exlstentlal tnformattbn for such

aurpos8s as MT,What a pureiYJlnguisttO notatton Will almost certainly no; be able

L- to do is to Cope with complex inferences of truths from other---the

truihsPurpose

all,for which the predicate oaloulus notation va$J afte;

-. withdy;vl&d,But that wIlI not be so great a loss when we rre deal!np

text of any degree of soohistfcation and oomDledtY for.translation,For in ihe world of real Words, and outside the wo;lds ofblocks and steeples, the kind of Inferenoes that a banaus’fc loglo oi

c cornron sense statements offers will not be of much use,

Let me give an exam~(8 of about infarences,and from a I\n9utstlc

bb

SOUrGe, In a rWef?t paperr Rlerwfscht23 sws that an adequatesemantics must expl fcate how "Many of the students were unable toangWr Your question qv fo(lo~s from 'Only a few students gfasped YOU;ouestion" Now, in a autte clear sense It do8sn't follow ai all;inthat th8r6 Is no problem about oonslderlng student3 who fall to graspbut nonetheless answer/That sftuation should not trsi anyone'sconcePtual powers very far, so It cannot be the case that on! follewsfror the other in the sense that if the premise Is true then theconclusfon cannot be false,We cou I d cal I that rslatlonshto otProPOSltionsthe:

Vhllosophlcal entailment", and 1 do not want iastatus of the notion

defendhere, but only to point ou$ that any

rep-resentation of the sentences in queatlon, logfoal or llngulat?o,that allows inferences like that one 1s going to be pretty useless,

There pay Indeed be a sense of "answ8r"f fn whloh the n&mL- Vx,~y,OUESTION(x),HUMAN(y),ANSWERS(y~ x)>GRASPSW, x1 would be a good

one t0 apply, in the sense Of aroduolng a true reSUlt, SUM there ateobviously senses of "answer" In whloh that Is Just not 30~ and toLpoint that out Is torepresentation,

demand, from the proponents of only logtcalsome suggestion as to how to cope with the real wo;ds

people us84 and to ask them to oonsfder that perhaps real language isk not Just an EXTENSION of dlscusslons of ooloured blocks,

e

8

1,2)the structure Of ihe translation and organkation sYstrm

The diagram below 1s intended to r@prOSeflt the WOrall rt;uotUre Oithe system under constructlon,

direct input of axioms in PC notation4

I lolrtcal II rePtesentatton I

IsemantIc**~***w*- irepresentatlonl -L**~******

I I 3 I I I IIENGLISH b-+-+*I I IFRENCH II TEXT 1 cc-4--cc 1 I*---ITEXT I

4 **a--a-w******** 5

I-nputt 4 9 4 4oara- system semantto sY8trm TRANSlaiacl outautgraphs axloms dtctlonary axloml, ParaaraPhs

entrles

1 assume In what follows that DrOGesses 2, 4 Wtd 5 are the ;alatfvrlyeasy tasks ----ln that they involve throwing awaY inCormafi0n----whlls1 and 3 are the harder tasks ln that they Involve maklng lnformatlonexollclt with the aid of dictionark! and rules,

With all t h e p a r t s to the dfagram a n d the facflkbs they lrnply---Including not only translation of small texts vla a srmanGicreoresentatlon but also the tranalatlon of axlams In t h e Predicatecalculus CPCI Into both natural languages -a--a-- i t ‘fs clear thaglnpu:t to the system wst be Pretty much rorrtrlcted If anytb'fng Ia t0be dono In a finlie time, However, there arc) ~i#arly W4Ys ofrestricting Input that would Just destroy the point of ihe wholeactlvitytfor example if we restrIcted ourselves to the tianslaiion O#Isolated sentenoes ;aiher than going for the t;anslatfon of da;agraohlength texta,Whatever Bar-Hillel says to the oontraiy about MT beingossentlally concerned with utterances Cl?, I am asrumlng ihat $hsonly Sort Of MT that will impress a dlslnte~ested Ob8lrp; will bethe translation of tsxt,In any ease ooncsntration on utterancea oaneast IY lead to what 1s In fact conaenttatlon on the t&k sxamdlesentenoas of Ilngulsttc text books,

e

L

.-

1

LIL.

I

L

So khat is to be the general strategy of tranalatfon? IZ f8 tosecmm the text in some acceptable wayI produce a ssmanFtcreeresentatlon as directly as possible, and generate an ouiout Frenchtorrr tram It,Thfs would Involve mapolngtemplates directly onto the

what I o&Ii 9amanGlaclauses and phrases of Enollsh, and

trying to mao out directly from t‘he tewlates into French clauses sndphrases, though with their relative order changed where1 assufre a l s o , that no

necrassaty,strong syntax ana/Ysls, In the lhgulstic

sense, Is necessary for this DU~DOW and that all that la necessarycan be doile with a good semantic reoresentailon----Wh\ch leaves USwith the big questIon of YJhat Is in the semantfc box, and how Is rtdifferent from what 1s in the logic box?

In the dlagramr I am using "samantlo representatfon" narrowly 40 meanwhatever degree of representation 1s necessary for Mftnot neoessarflyfor auestlon answering t that’s what the lo910 box ts to;) or fort h e o r i e s o f how t h e brain w o r k s - - - - - a s little representation as wecan - get away with In fact -----MOch I am personally sure fs horn ghebra1n really works, For this We maY welt not need the reflnemants otV 1 g 91 that I mentioned earlier, norr say, exlstantlal quantffTcrtlonor the ana!Ysfs of oresuPDoslt1ons given by translation oi dsflnltsdesCrlDtlOnS, MY naln assumwlon here about the difference betweenthe two boxes, logical and Ilngulstk, fs that an "adeouateet 'loglca1translation makes all juch matters explicit, and that is why ;i 1s somuch more difficult to translate into the too box than the bot;ombms,But the dIfferem@ between the two remainsone;intended to cor;esoond to two "levels of

Drawnahunderstandl~g~~ In the

human befng,

with the dlfflcult t a s k i aohleved, translatlon f;om seman{icrenresentatlon Into a logical one, then lt might be POs8fble to havethe two paths of translation from Enallsh to French;namely 3-5 and3-l-2-%Jhe translation through the laafc and out again might not beeSPeclaf IY illuminating but ft would b9 a aQntrOl Chat $hOUld notproduce a noticeably worse .translation than one achtrvcld b y ;hesJ70rtW route,

Inouts t o the logic box wll I be In a RestrIcted Formal Language(RFL)Cse% 51 and It should be ~osslble to Input ax/ems ft~ ii dlreciat a-screen or teletYpe,fhe RFL will have to be at least.as fo;mal agthe @scriptIon 1n McCarthy and HaYsstt3J 1f the diagram ls to be ofany 'USB# for the;e 1s no wlnt In having an RFL ta ENGLISHtranslation routine If the HFL 1s close to Engltsh ----one might Just:a s kell Write h English, The Sandawall QormClBJ, for examdIe, withfnflxed Predicate names 1s probably already too like Engl~sh'Jha?sno argument against hfs notation, of aourser simply an argument that1t rrlght not be worth wrltlng a translator from 1t to English,

The nature of the mapD!ng down from loglo to the Iinoulsticreoresentatlon WI o f cOUrN3 dsoend on the relative sizes of {heb~nventorIgs ot primftlves and form In each! however, one may exoactthat the tIeId of logloalpr1m1tlve Predicates dill be a IarQer one

and that the maonlng down will be many-one---- with a number oflogical exPressIOns mapplng onto a slngle semantic temPlat8,

at It should turn out that the level of underaianding Provided bY $hssetnanth coding ls inadequate for MT, then the dlawam can stl I Iaw I Y to the logic box funotlonlng as the lntarlln9ur~ths djfferenosbeing that the semantics ~111 then be effectfvely a iransladon stagebetween natural language Input and the looloal representatfon,

If the gemantIc cOdln!J does turn out to be adequate for sbme form ofiestrloted MT then ihe tunotlon of the loslc box will be In ihsanworlng o f auesttons abouttranslated,In that cage only those

the Content of what ha8 bpnstatements from t h e ttanslated

text relevant to the questton need be translated UP fntb the lo&fOirr,,

What follows Is dlvfded Into four Parts whloh oorresoond tb sta9Os onthe diagram above,

2,l)The processI?ig o f Enollrrh Input text, 2,217he intrrlin9ualreprsentation Produced, 2,S)The form of the dlotfonary Ured', 2,4IthegeneratIon of French Output from the tnterlln9ual roRr~8oniaifOn~

2,i)The Orooesslng of English text,

The alm of the taxi prooesslng secttons of the overall &og;am fs toderive from an English text an tnterllngual rsOresentat~on <hat haaan adesuate, though not exoesalve, OOmPlexltY for CWO tasks:

f)as a repr@sentaitgn from whbh outout In another natural language----French In this case----can be oomPuted, 1bas a rap;aasntatlonthat oan also serve as a n analysavdum of Predloatr calOulusstatements about some partlculap universe,

The flrgt pass made of the Engllah tnput text la the fragmentationand reordering procedure, whose functton is tO OartltlOn and repaOktexts of some Iength$U sentential complexity Into the iorm mostsuItable for matching wlth the template forms mentioned above,Thlastag-8 Is neoessary becau88, like all proPosed coding schsmos, loQ!oal

Ilngulatlo or Whatever, the template format 18 a more Or I@SB rlgldtne and the awful variety of natural language must be made ib !lt, If'the system Is to analyse anYthIn more than simple example stdencos,

As I mentloned earlier the baato format of a templaie Isa8ubJ~ct-verb-obJect one---o tn

------!~u~h a8DUrQlY goman+ ie-m8#

actor-act-obJect one M A N HAVE TW’JG, w h i c h WOU~:hopefully be matched as the bare template name Of any sonlencr 8uOh

"John owns a oarO,MAN, HAVE and THING are ~nterllnoual elsmen?ls,lid MAN for example would be expeoted to be the arlncloal, or head,element for any semantic formula reprerentlng ihe fnallsh wo;d tfJohnvQin the dlctlonary,Sfmflartyr HAVE would be the head elemeni In 6heaoProPrlate semantio formula for "owns? and 80 on, A slmdle

cmatching algorithm would then be able to match tho acceptableseouence of head elements MAN HAVE THING, which is already known tobe a terrplate, onto a sequence of formulas drawn from tho d1ciionar'yfor the word8 of "John owns a cart”,

c

L'

The details of the maiohlng algor‘ithm are not of concern herat whai1s Mortant to see ts that an algorithm for matohing a ba;@ th;eemelevent template onto a piece of language by insoec$fna just ihehead elements of formulas and searching for acceptable;theff,

80auenoes ofwill, In the course of maklng the matoh, select no; only {he

head element of the Word formula, but with it ihe wholewhich

?oimula ofit was the head, where "whole formula tt Is to be undeisiood ai

this Point as a coded form that BxPres3es the whole content bf Gheword Sense in ouestion, In the Present case rvJohnvlr bejng a merename, has no sense Over and above that It refers to a human being,and its whole formula would be simple (THIS MAN) WhiOh says no mOrethan that,

One of the hypotheses at Work her8 Is that there Is a flnfte‘c- inventory of t'emplates adequate for the analysis of ofdinary

langua90--a usuable I'fst of the messages that rreople wani i;o conveyI with ordinary, language--and that ln selecting th088 sequences of

i forruks for a fra9ment that are also t e m p l a t e sequenoeg ( a@ iega&their head elements ) we pick UP the formulas correspbndfng to <he

i

CORRECT, aPProPrlate, senses of the words of the fragment, as theya r e belna used in that oartlcular fragment, I am otvino only ahl9hlY Qeneral d@sOrlptlOn here, and the details ofof this method of

the aOolfcationanalysis to comPllcated text has been set, out In I:

I 153,L

L

L

Moreover 8 It is supposed that any frasment of natural IanguaQe canbe n a m e d b y , that 1stemplate c

to say matched with, at least on! such bareand that the name will $erve as a basic

fo; the PurPose of translatlng the fragment,core of meanlng

Or In othei wo;ds, wecan know how to translate from the oomnPlex+3resw7tatlon

tntsrllngualof which the bare template MAN HAVE THING 1s the name

simplY becauroe We knaw and can reduce to algorithms how to exbressthe message "a person has a thing" In French,The template ia thus anhem, or unit, of meaning to be translated,

An sx~mple might help at this Dolnt to gfve the general ldrra of whaities are establishsd between text Items bY the matching routines Ihave descr 1 bed, Supbose we apply the template matohingthe

iout’9na t osentence: "MY brother OWr3S a large 0aP And lot US

SUPPOSE furthermore that we are not ooncerned with the probl8m ofselecting the CORRECT sense formulas, one corresponding co eaoh ofihe words in that sentence, as it is used in that sentence,We ahal Imake the simRlifYing assumption that each of those six Woids has onlyone sense entrY in ihe dictfonary ) and that what we are consideringare the relationships set UP IndirectlY among the words by matchingan interlinWal rePresentat!Pn onto the sentence,

12

From the point of vtew of the matchIn routine, the lnltlalreprrsentatlon of the sentence Is a string of six ssmantlo formulas,whose detalls I shall discuss latercAt the moment what matter8 isthat the formula for vwbrother"the one for

has the head element MAN@. Just as dfd"John", and SO on for trowns~' and "oar", The formulas t.0 r

0,y " and "larger9 have the oonventlonal head element KIND I Sin08

thry soeolf~ what kind of thfng Is t n gurstlon, The tomPI atematching routtne soans the formula string from left to right and isable t0 matoh the bare template MAN HAVE THING from ihs ternolaf;Inventory onto ihe f o r m u l a s for "brother" r~owns" and "oar"rssDsctlWY, slnoe those elements, In that order, are the heads ofihose formulas, Those three words ap, as It wsrel ihs dolnts Iniho sentence at which the template puts Its three feet down,

SO far, at the word level I tles that oan be written as f~llb~r h a v ebeen establIshed

brother 0 owns @ ca;

Those aye much thg same sort of tbs that would be establfshrd AT THEWORD level by any system of conoet2tual semrntlo anrlyslsCof,~l3applied to that sentenoe,

T.hts word dependency thenr Is set up by matohlng the bar@ iemplatr ofelem8nts MAN HAVE: THING onto the string of foimutas for the Kurds ofthe sentonce, This In Itself Is no vaouous exerolse beoause, givenChat all reallstjcall~ coded words In the dlcclonary would haye manysense formulas attached to them, only oertaln sslsotfons o? tarmulaewould admlt of being matohed by an Item In the template !nventorY,Forexample, In the sentence *tThls green bicycle 18 a winner", sheaemantfc formula for 'Wnner*' that has MAN as its head and means %nswho wins" Is never ploked UP bY the matching rouklne stmp’iu beoausethere Is no bare template THING BE MAN In the !nventorY,

7-o return t0 the sentenoe '*MY brother owns a large oa;V hrvlngmatched on the bare template, the system looks at the thiee formulasft has so tied iogsthsr by means of thdr heads to gee If it Ganextend the representation, top-down, bu attaching other fo;mulas and

-create a&mYla for

fuller rePresentatlOn' In this oaae it looks f;om the“brothe;” to the one that pryceded It

fortriuls for “mY”,fhfs It seesnamslv $he

oan Indeed aunlh &r &mula fo;@'brothePrihat

and so It opens a (1st of formulas that can be il,ed ontb"brother" fOrmuIa.

IntsrlMual rePresentatlonR;;;atl;i this Drooes;fWet;;d up with an

sentenoo tollowlngschewatlc form (that I: shall call a FULL TEMPLATE----though we shallsee later that the iled items ars not simply formulas )I

FCbrotherJ FCownsl Fkar)

(FCmyJ) ( FClargel)

13

where both the hotlzontal and vertloal dtrectlona boresentdependency tkl of the sort I have described and FCxl atmply standsfor the interllngual formula for the English word x,Thua thevertical

uwa;dsdependency

in the case ofIa that of a-3lst of cauallfylng foimulas (amoty

rrOwngtt) on a malt7 formula,

The correspondfng tiea between the text words ihemsslveg astabllshedby this method are:

brother - owns * car * a? ?

my large

A pofnt that cannot have escaped any reader Is that by having 8 rigtdactor-actlon-object COrmat for templates, on? ignores Chr jacrt that;;ny$agrnent? of natural language are not of thla form, bgardlrss

the Intt_f,al InpUt text Is Prrtkloned, This fs indeed iheease, but p as I shall describe, by ustng the notlon of dummy DGtsof temPlatea one can in fact Put any text oonstructlon into th?s verygeneral format, Since the analysis has no oonventlonal syntaohbaser the standard e x a m p l e s of syntactic homonymity, such a s theVadoUs tnterpretatjops that can be thought up fOr f’they a;s eatingaep 18av', are raPresented only astnterPretatlons,So, for that sentence we Would

dfffertng measasmexpect t o matoh ai

least the bars templates MAN DO THING and THING BE THIN&

FRAGMENT AND ISOLATE

The fragmentation routlns wwtlttons hut sentences rt pUcclJdltT0nmarks and at the occurrence of any of an extsnslve, though tlnftelist of key Words, Thatsubjunotions,

list t h a t oontalns Cglmost ~1;conJunctIons and preposttlons,

1s in the house@@Thus the sentence "John

(J6hn 1s) andwould be returned bY auoh a routlne a8 iwb #ragm@nts

(ln the house),Wlth the first fragment the sysgem wouldmatch MAN BE DTHIS, where the D of DWIS lndfcates thai, havingfatled to find any predicate aftor 0ls't, the ayaiem h&sdummy -ThIS to Produce the canonical form of template,

suoP!led a

When II t comes to choosing the cOrr@ot template for the fiagmsnt, Ifthere la more than one available to choose from, the general overal Irule Of chofue that I ref8rred to earlier, of always prefeirlng ihereprssentatlon with the r*rmt conceptual connextonstwhl~h oan be,thought of stmrw as the number of +' s In the word dlagiams), willd;aYS choose one WIthout a dummy ln Dreference to one with Though ln

Present case only a template with a dummy would be ava labI@ fortchoosing, In the case of "In the house" the maiohlng rougjne findsitself oonfronted wtth a string of formulas starting wfih one forWVt that has PDO as Its head,Preoosltlons arer iP generalasslfll lated to actions and SO have the P In the PDO of the]; heads t;distinsulsh them from straightforward actlon formulas,the

I n Chia CasematchIns routb Inserta a dummy THIS as ihe left-moag member of

14

Che bare template I shoe It first enoount8ra aotlonformula--headed by a PDO--a8 It soans the formula strininfrbm lrfito right, and, ‘4 n th8 houa8" Is fInally mafohed wfih ih# b a r etemplate OTHIS PO0 POINT, So then tho sentenoe "Jbhn 1s ln ihehougetf Is PartItIoned Into two fragments and matohed wlth a srmantloreprssentatlon consisting of a string of two t8malates who88 baretemplate names are MAN BE DTHIS and DTHIS PO0 POINT, resdeotlvsly,

Another examole of fragmenting and m a t c h i n g 1s prasenied by whaimlght conventionally ba cralled noun Phrases, If R after Cagm8nttng,the system Is Presented with "The old blaok man" a8 a single fk#mrniiIt SUPPlY two suoh dummies during the m tch and end UP with a;ror%ntat!on named#by the bar8 ternplate MAN DIE DTHIS,

The remantlo oonneotlvtttes d8scrlb8d so farr 8)ementary though they

tre 8 have been between formulas that corr8sPond to wotdrr ocourrlne

the samef:rQ*entatlon

hOrnant Of text@ The gr8ai advantage o# {heapproach Is that It breaks a sentenoe of ) ParhaPs,

thirty words, ln.to a number of units of manageableoomPlexlty,

lntqaland such that a template oan b8 matohed onto 8aoh In the

manner d8sorlbed,

But not all semantic ties In such a oompl8% rsentenoe will be Internalt.0 fragments--many will .be between ttems oocurring In dlffsi8nt,and maybe not even textual IY oontiguaus, fraQmon&,At a la-e; doin{ Ishall discuss TIE iouChes whose function 18 io ProVld8, 4 n the fullInterlIngual reoresentation, those Inter-fragment dgdsndeno!rsnec88sarY for translatfon,However, thet;$or s$$fYlng role of thefragmentation must not be lost In all Is to allow aromp 18% sentenoe to be represented by a Ifha sequenoe ol {ernplaeeswith tlas between them---rather than by a far mob oomplexhlerarohloal represeniatlon as Is usual In llngufstlos,

the fragmentation, than, Is done on the bath of th@ aude;ficlalpunctuation of th8 Inout text and a flnlte list of keYWo;ds andkeyword sequences, whose ocoUrreno0 DrOduc~S a to%; daihbn.DifflcUlt but lmooriant oases of two kinds must then be trlrstlu,

eons der8dlthose where a text string Is NOT fragmented oven though a

key lword IS encountered, Two tntuttlv8lY obvfous casSc)s are nbn-subordlnatlne use9 of Vhat" as In "1 llke that win8f'r andprepidton8 funutl0ntng as “PO3 t vorb!P a8 In "He gave UD hfaooat',In these oases there would be no fragmantaCfon br&o ih8 keywords In other cases text strings ape fragmented even ihbugh a keyword IS NOT present, Four oases are worth menilonlnor

I )“I want him to go" 1s fragmented as (I want)(hlm to go)iA boundaryIs Inserted after any forms of the words **say** and "want% and afuither boundary fs fnhlblted before the following $orct Thts~OWV,S lntultiv8lY acoeptable slnoe rcwanttV tn faot subjofns the who I,?Of what follows 'ft In that sentenQe,We shall expect io matoh ontoGhose fragments bare iemolates of the form MAN WANT DTHIS and MANMOVE DTHIS rrsPeotlvely ---where the first dummy THIS fn faot stand8

foi the whole of the next template, The fiagmsntatton funotlbn8OPerate a t the 1 OW89t oOsslbl@ leV@1 of anabds, whfth Is to saythw fnspect the semantfc formulasdictfonary,

fJlr(@n for 8 Word I n thebut they cannot assUm@ that the chotoe among the formulas

hag been made,

So then8 the fragmentation functlons can canslder only the Tans8 ofPOSSIBLE senses of a word,Howev@r, In this ca88 InsDeotfon of any o f

"says" enables the system to ?ni@f thatact cm subJoin a whole template and not merely an object4 as in

“1 lllrant hlm*c,A verb Ilk@ QdvfSe" on the Oiher hand IS no; Of thfSsort since we can tnfer "1 advfse hfm"want hfrr** In the ea;ller c a s e ,

ln a way we CANNUT infer "ISo we would expect o I advise him to

go** to receive n0 spedal treatment and to be fragmented ag (I adVrgehimHt0 901, on a key word basis,

IIIRelatlve clauses beglnnlng with *'that'* or "whl oh'* ar8 located andisolated and then Inserted back Into the string of fiagm8ngs ai a n8Wpotnt,For examDl6-."The girl that I lfke l@ft” ts fragmentedSW l@ftUthat I

(TheIIke PD)iwhere the fInal D8rfod of the ifntence

tt p 0 ” is alSO moved to close off the 88nt8nOe at a newthe partftlon after “1 Ike**

Polni, ThusIs made lrl the absenoe of any key w&d,

III)“The old man fn the corner left*' is naturally enough f;agmrn{edas (The old man)Un the corner)(l@ft),Th@ breachthe and act

m a d e hei b8tW88nactor of the sentence Is replaced later by a t?e bee

betowL

L

L

Iv)The sentences “John Ilkes eatfng flshfg “John likes eating” “Johnb e g a n eating ffsh (( are at I fragmented before "@atlng", so ihai th@s@forms are all asslmflated to **John likes tosynonYmous with

@at f i s h 'r(wh!oh tsthe first sentence

eailng f fsh**,abWe)rath@r than i0 "John IS

"Johnwhich would not be fragmented at all,In tempiaie

Is eating fish"tGms

"John likes fish"Is to be thought of as MAN 00 THING, while

fg MAN FEEL DTHIS + OTHIS 00 THING, Where ih8 fi;SiOTHIS refers to the whole of the next ternPlater and the secrond DTHISstands In Place of MAN (!,e, John),

**Oft* 1~ a k@Y word that reaelves rather sp@cfal treatmen& and la notused tb make a Partition When ft Intfoduoes a possesslva noun dhrase,After'fra9mentat~on

iunction, each fragment IS Da898d

whichthrough an

wlthlnISOLATE

looks e a c h fram8nt and seeks fo; ihe righihana boundarfes of '*of" phrases and marks them off byc h a r a c t e r tnto

insrrt"lng aI’ F 0 ” th8 text,Thus **He has a book of rntnvf would be

returned from the ISOLATE functfon as **He has a book of mine fo*q,Thfs1s done In all cases except those Ilke “I d0nt want to sDsak of him"where "of '* effectively function8 as a post verb,

It may geem obvious enough why *'of" Dhrases should remain withln ?,hef ragpent sincedemarcatfo: of the

"0 9 John:, functions as does '*John's", but <heDhrase with the 0 F 0 t' character

explaIned by considerbvg the PICKUP and EXTEND routines,oan only be

PICKUP AND EXTEND

The PICKUP routines have alradv been described In a Wnefal way tthey matoh bare ternplateS onto the 8trlng of formulas for a textfragment,As the routines move through the string of formular, thosecontained between an 0): and a FO are IOnorbd for the purdous of iholnftlal matoh,thls ensures that “of p)lra889 are only $;aaied a squallftrrs,So, In ihe sentence Vhe father of my frfend f$ 1s orl[edJar;kvqr the match would never try to make the head of the formu a $0;ttf r I onndrr Into the ;oot of a temnl:ate matohlng the sontwo+, s no8 ittfs sealed between an ~~oj--f~@@ Dalr, T o lilusirate the rerults oirprMlnQ PICKUP, 1 shall set down the bare ismplaiea {hai would beexeected to matoh on& Nlda 8 Taber'sC83 suggested sov@n basic fo;7mrof the English IndfcatfVe sentence,(In this not8 I ds@o*\be only ghe'tndlcatiVe mood as lt Is Implemented In the tFla[ vet8 on of Chjs68y8tom,Quetlas and lmperatlvee, like passives, aro deal; wlih by GheapprbPrlat@ man!oulatfon of the template order,)

In oath oaee I olv-B the baa/c sentenoe, the brie templaie, and adlaQramatlc repreeentatlon of the oorrospondln~ dependen~le~ jmPl\edbetween the tewt items, where oav( agrln Ilnks those words bn whlohihe bare template 1s rooted or based, and N-V links a dependent wbrdto Its 9ov@rnor,

i-

‘i

I-i

c

L

L

1) John ran aulcklyhAN MOVE DTHIS

John * ran - CDTHIS]?

quickly

illJohn hit BillMAN DO MAN

John - hit - Etll

iff)John gave Bllt a haltMAN GIVE WING

John - gave - ball

(to&I :--,The Wtabt fshment Of thfs dwmdenoy by EXTEND I S discussed below,

lv)John IS in the house,MAN BE DTHIS OrHIS PBE THING

John - 1s w COTHIS CDTHIS3 - In - house9

the

v)John is sickMAN BE KIND

JQhR - is - sfck

viUohn fs R boyMAN BE MAN

John L Is - boy9a

vii)John is my fathe?MAN BE MAN

John * Is w father9

mY

6

. .

A natural auestlon at this point Is what exaotlY Is thk fnvrn~orY ofbare terndates to be used In the analysis of Input lanauaae? Nb -*detailed defense 'fs offered of the inventory used,nor,l kelfave oanone be glven,The fact IS that one U$@S the Inventory that seclmgemplrlaaW dOhtnrevIses It when. neoessary&l opera@on oi undercrltlclsm,and concludes that that,alas& how things must be In $hereal world of practfcal IanWaQe analysts,

Ths InventorY used can be reconstructed from the table of des eegOIJ~: below In Baokus Norma) Form,It 1s set out In terms o? the aodondeslgnatlng semantlo elements,such as FORCLand the olasses o frwbstanttve deslgnailng $ elements (such a8 *SOFT maan ngSfUFF,WHOLf,PART,GRAIN AN0 SPREAD) that oan Preoede puch an ac~lo;aq:a subJect #and follbw it as an obJect to oreate a t h r e e elementBernplate,

<be@ tempWe>l:o<*p(p<Do><~W I<rPO><CAUSE><*EN>k-.<*PO>CCHANGE><*EN>I<*AN><FEELW*MA>I<+EN><HAVE>C~EN>l(#AL><PLEASE><+AN>(<*AL><PAIR><+EN))<#'O>CSENSE>C@EN>I<*PO><WANT><*EN>(<*POMUSE><+EN>I<U'O>CTELL><+MA>IQ*PO><DRDP><GN>IQ~PO>CFORCE><+EN>IQ+EN><MOVE>CDTHIS>)<*PO><GIVE><*EWC+AL><WRAP><+EN>(<*AN><THINK><+MA>Jc~SO><FLOw><OTHfS>l(*PO>CPICK><dN>IQ*PO><MAKE><+EN)J<*AL><8E><ram@ mombe) of #AL as last oocUrrenoe>

<~AL'j~~~<~T~~S~TH~S~MAN~FOLK~GRAINl~ARTl~GRL~~STUFF~T~ING~8EAST~PLANTlSPREADlLINE~ACTlSTATE>(*AL moans all aubstantlve element8)

<~~~>~r~<~T~~S(tHfJ/MANJFOLKIGRAINlPARTlSTUFF~THING~8EAST~~LANT~SPREAOiLINb(*EN moan8 el$m@nts that are entitles)

c~AN>lrr<MANIFOLKlBEASTlGRAIN>(*AN mean8 anlrnate 8ntltks,CRAIN 18 ua@d as the mainelement foi 8oclal organlzatbnsdlke The Red Cross)

~~p~,rrs<~T~IS~tHfS~~ANJFOLK(GRAI~l~ART(STUFF~THI~G~ACTi~EAST~PLANTISTATE)(*PO means potent dsmwtsdhose that oan deelqnatewac$orstThe olaes oannoi: bo r@rtrkted to *AN shoe raln w@ta th@

19and the Wind opens doors)

<*SO>:t=<ST'JFFlPARTlGRAINlSPREAD)<*MA>:!=<ACflSIGNlSTATE>

(*MA designates vark elemsnts #those that can dsgtgnataitems that themselves deslgnate like thoughts and w;ltinsW

It kill be noticed that I have dfstorted BNf very slightly so as towt 1 te the bare templates oontainlns BE in a convsnlent andD~rSDiCuOUS formqb forms containing MOVE and FLOW also

(0, they apeconiatn

DTHIS @'dummy tem~latW*) indlcatfng that there oannotbe oblects in those bare templates,Thus MOVE i9 used only In {hecOdh3 Of htransitbe actdOns and not to deal wfth sentences like qmoved all the furniiure round the room",

There are dummy teqnplates not included In this Ijgt ---ggvg;a( a~&L i n t h e dW3crtPtton o f the Nida and Tab& sentences above, The

remaining rules specifying them are intuitively o b v i o u s , & may befound* in detail Jn Cl5304here I also glw important anallliary rules

b- wh t ch SDWlfY when dummies are t o be generated ln matthfngsentences,Naturail~ a, dummy MAY BE OTHIS is generated lo; ih8 fiist

Ifragpent o f (John IdUn t h e hoWe) slmplY becaUSeelement

PrOPer threebare tamPlate cannot be fitted on to the ?n?oGmatlon

availabie,But in other cases,where a thrgg elementfStted,

templab oan b e

i

dummies are generated as well,slnce SUbSequent routines to bedescribed rrlaY want to prefer the dummy to the bareexamle

tsmplate,Fo~in the analysis of the first fragment of (The old transpbrt

I

sYstemHw;ich I loved >(fn my youth)(has been found uneconomlo),areasonably full dlctlonarY will contain formulas for the substantive

L sense of "0 1 d " and the action sense of "transpor~",Thus,Ehsactor-action-object jgmnlatg FOLK CAUSE GRAIN can be fftied on herebut hi!1 be incorrect,The dummy GRAIN D8E OTQS will also be fitF,edon and WI I iL be preferred by the EXTEND procedures 1 describebelow,Such slight oomplexitY of the basic template noZIOn areneceasarY If so gimPl8language,This matter

a concept Is to deal wTth the reali;Cies ofis described in greater dgtall ln [153,

.The rratching by PICKUP wfll still, in general, leave a number of bare

i templates a t t a c h e d to a text fravwnt, It is the EXTENT foutlnes,I workfqg out from the three points at which the bare template attaohes

to the fragment, that try to create the densest dependency networkoossble for the fragment, In the way I descrtbed earli& and SO t oreduce the number of templates matching a fragment, down io one i fpossible,

In order to show more clearly how EXTEND does ihlgr ft fs necessaryto 5aY somewhat inore about the semantic formulas Which make UD Ghefull tefrolate, A semantic formula expresses the 3eaning of one senseof a r?atUra I language word in the dlctlonary,lt is made UD of leftand right Parentheses and of semantic elemenis,the latter lnoludeTHI&Gp STUFF, MAN etc, far bask Items in the world;FORCE, CAUSE,DROP, Ch4NGE to describe basic ktnds of action, and SO on,Thef o r m u l a s are bfnarlly b r a c k e t e d pairs of dhatever d e p t h of nesting is

2 0

necessary to express the meaning of a oartloular word sense, Theformulas are made IJD~ and Interpreted, with a drpendenoy bf the leftelefr,ent, or bracket giouo, upon the oorresoondfng rlaht hand slemsni0r bracket group In every oabe,

SO8 (MAN KIND) would be lnterprotedh "of a human sorV, which 1s tosay 8 It Is a formula for rfhuman" used as a ClU4l ifi& Xnf(MANDROP)CAUSE) the dePendenoY wlthfn the Inner bracket is ot anaCtOr"aCt type, whereas that wlthln the outer braoket sD"Cll"-mof (HANDO) on CAUSE-- Is of the objeot-of- actlOn on aot tYPe,So iho whbleSub' formula person to renounoesomethIng%

lS t0 be interpreted gS “0gUSeSand we would therefore eW@ct to find th?s subformula

within any f o r m u l a to; ( saYI "blaokmali" There a r e ~~&Toiibnfi

on the ways In whloh the elements can oom6lne contaIned In a table oi!vcope notes" for the system of oodfnglfor e%ample, CAUSE cannot beanything but an a&on so ((MAN QROPXAUSE) o o u l d nb< .be $hespeclflcrtlon O f a sort Of oause but only t h e causfnu oisomethlng,The most important element'ln a hrmula Is Its ;fghtm&iono, 0r head w i t h - which FICKUP oonnecta formula8 fbr WGd@ tbiamplates to r whole fragments In the way I descr/brd,

Fo;mular that c a n qualify any #other, substantlve formula hrvlr ihehord KIND@ and those that can ciuallfY actions have the head HOW,Mbsiroil~n formulas hava IRS head DO, BE8 MOVE ("run" for eiamp 011 0;(#&GIVE verbs are Important In that they oan function 4 ihereprrsentatlon of actIon oonstruotlons Iike "He left John is w:tch"where an lndlroct obJoct of an aotlon can appear w ihout an;7oreoodlng PrePos~tlon,GIVE verbs functfon l'n much the same way asTRANS verbs In Sohank's analyslsCliJ,and the a0Daarance of GIVE as afoimula head for , s a y , the aotlon@~left7~ primes the s~sbn io expeo):auah an Indirect obJeot,Ths verb "tell" also has GIVE as the head ofIts Prlnolpal formula ahoe it oan Partiolpata in suoh lndlreo<obJect oonstructtons as nJohn,tells me a story",the lack of neo@ssaryaonne~lon betwesn the English word ‘WI In and the !nierllnguale'len;ent TELL 1s brought out by this faot that the f0rITIula hrlad oi"t 0 I I " Is not TE L,the head of its ma n formula 18 TELL sinoo lt oannott

but CIVE,fn the cam of resay’* on the other handOOCUr In ihe

CIVf+Yp@ COnstrUotlons,

Most: substantive f0;mulas have as their heads suoh elements 8s MAN,STUFF I THINGI ACT(for abstra0t aubrtantfves whloh ape the fesult of8ctlan, such a8 "adJustment% STATE (abstract substantives such es@@frlrndshWr f~hapotnessV, GRAINhbstraot substantlvea any sort ofstruoture suoh as "systgm") @nd 90 on,A formula for

t Pau stantlve 1s

assumed t0 be singular unless the e(c)ment MUCH la ts li;~i Item al,ihr to0 level,

Aoilon formulas can apeclfy a pref@ri@d olas~ of arotors 6; of obJeatSof the aotlon or b0th,Prrferred rotors ah spc)o/fi#d bY FOR andpreferred obJects by TO,So then the formula for the aotlon ‘1 i a I k '9wfll eontaln the pair -(MAN FOR) alnor most things $hat talk a;~humant and if there is a posdblllty of uottine up Eli depmdenoy with

21

Q

c

c-

c

b

c

b-

L

a human actor, the system will take It,The restrIction cannot beabsolute In this, 0; most other, oases since machines and dogs talkin fab I e if not In fact, The Important facflfty is to be able tkPREFER the usual If a representation for It Is available, but to beable to accept thk unusual if neaes-SarY,

The syntax of the actlon formula is a s follows t(X FOR)or (X TO)appear a8 the first item at the too IWei of the aotion foimula ffthw are appr0Prlate -m,-,;n LISP terminology the pali 1s sfm&CONSd onto the verb formula, If bath are appropriate, as in aformula for ,,interrogate", then the (X TO), for the obJects, 1s CoNSdfirst, and appears at one level lower in the nesting of the formulat h a n the (X FOR), speolfying the Preferred actors,Thus ihs formulafor Vnterrogate,, would readt((MAN FQR)((MAN TOHTELL iVRCE)H, Thepreferred substanihes ,or c l a s s e s o f them,for qUrllff8iS atefndlcated naturally in an extension of this notatfon ,by including (xFOR) as the f I rS* item a t the top I eve I In the fo;muln fo; aaualifier,

In order to ykeep a small usuable set of lntarlingua'l ssmnntftelements,and t o avoid arbitrary extensions of the list ofelefr3nts,maY notlons are coded by conventional sub4ormulas;(FLOWS T U F F ) ia used to desfgnate ltaulds for examplerand (WHERE SPREAD) to.oode spatial area of any sort,

After this brief description of formulas, some further sp~olfloatloncan be given of ihe EXTEND routine, whloh Is absolutely central tothe a n a l y s i s # since :t is there that mosi: o f the work of aconventional syntax analysis is done by semantlo methods,

I explained the role of EXTENO In general terms earllet t It insaectsthe strings of formulas that replace a fragment, and seeks to s@t uodependencies of formUlaS upon each other. I t keeps a glcor@ a s Ii:does so, and in the end selects the structuring of formulae with ihe-most dependenotes, o n the assumptton that it Is the righi on@ co;3nes, if two or more structuring3 of fo;mulas h a v e ihs s a m edependency score)

The. dependancles that can be set UP are of two so&t A) thbsebet%aen formulas whose heads are part of the bare temolatsi @> thoseof forfru(as whose heads are not In the bare template upon thoseformulas whose heads are ln the bare template,

Consider the sentence "John talked ouickly,, for which ihe baretemplate would be MAN TELL DTWS, thus establishing the dependencyJohn - talked - cDTHIS] at the word level' Now 9uPpose we expand ouifro@ QCh of the elements ConstltUting the bare temRla& ‘fn &n,Weshall find that In ihe formula for ,,talked,' there IS the preferencefor an actor formula whose head Is MAN--- since talkins h generralludone bY people,Thls preference is satisfied here, whioh we can thinkof as establishing a word dependency of ,,John,, on ,~talked,,, whloh isB iy~e (A) dependency, Expanding again from the elemen; TELL we

22

have a f o r m u l a fOf "quickly" whose head 18 MOW, and HOW - h e a d e dfovrulas a r e p r o p e r quallfl@rs for actfons, ~enoe we have been ableto set up the followIns dlagramatjc deoendency at the word l@vdl

John - talked Y CDTHISJ.* 7 . .quMW

(where " - " Indicates a bare template connectkity strsng{hened by al

direct semantlo dependsncy~a~=sprlnQlng from the preterenoe of"talked" fo r a human actor In this case,) and we would soore two forsuch a repreaentatlon, Furthermor@ ? the formulas havjng type Bdependenoe would be tlsd In a lfst to the maIn formula on which theydeoend, The subtypes of dependenoe are as followsr

A) among the formulas whose heads oonstltuta the bare temilate

Upreferred subJects on aotlons"John ta I ked" --.1I)preferred obJeots of aotlons on aotlona"lntorrogated a prisoner"

B) of formula8 not oon8tltutlng bare templates on thouothat do

l)quallfIer8 of substantlvos on 8ubatantlves“fed door"tl~quallflera of action8 on aotlons"opened qulokly"tll~artloles on substantlve8"g book"iv> of,,,, fo phrageg on substantives"the hou8e of my father tonv)quallfIers of actjons on quallCler8 of substantlveu-Very muoh"vl)post verbs on aotIonsllg I ve up"VII-) lndlreot obJecis on aotlonrr ."gave John a,,,,,"viII)auxIllarle8 on aottonstlwas Oolng"fX)"tO" on lnflnltlve form of autlon,lit0 relax"

The searcheer for’ type B depondenolaa are all dlreoted ln the formulastring In an lntultfvely obvloua mannorl(I) goes leftwards onlyW1)9oe8 right and Iaft(1il)leftwards onIy:~Ivlleftwards onlyl(v)loftwards onlYI(vl)rlahtwrrds onlyI(vlf)rl.gh_~rdr-:_ -. - - -_-._-The purpose of the SOOO~O ofd

23cclear if we consider an example Of l3Wi): the%dlrect Object

e construction,Let us take the sentence "John gave Vary the book? ontoL- Khich the matching routine PICW~ wi 1 I have matched two Dare

templates as follows, since it has no reason to prefer one to theother:

eJohn gave PjarY the bookM A 1\1 GIVE PIANMAh' GIVE THING

L--

L

e-

EXTEILD now seeks for dependencies, and since the formu!a for !?gavVhas ro preferred actors or oojects, tne too bare temOlat8 cannot beextendec a t al I and SO scores zero,In the case of the lower b a r etempla te , then a T??aNS action can be exbanded b y any substantfveforKlila to its imnedlate right which is not already part of the bare

IS nottemplate,ASafn Vook" is qualified bY an article which factn o t i c e d by the t o o bare template, S O then, by EXTENDing ~8 haveestablished in the second case the following dependencies at the wordlevel and scored twc (of the *We dependencies),

John w gave w book?

Ya ry tkeL

Two scores higher than ZBrO and thg sesond representation i s. preferred,This is an application of the general rule referred to

L earlier as "pick up the mast connected representat;on from theLf ragrrent", I wrote earlier of the relation of '*JOhrP to @'taikedqV inthe sentence "John talked quic;<(y" as being e x p r e s s e d In the ful Itenplate as a relation oetween temolate items WAN and TELL beingthe! heads) of mutual dependency #and SO not really a dependency a tal I, but strengthemd in this case by a "senant i c dependency" s i nce

c MA& is a preferred subject head for TELL verbs,But this form ofL exoression can be misleading becauss,in this system ,there is no real

syntax-semantics distinctlon at all,Every dependency is expressed byrelation5 of a single tYFe between elements and formulas and classes

- of both,even though some such relations (like the MAN/TELL one above)L clearly have a more semanticky f lavor,whi le tbse I ike the

b any-cubstantive/KIND relation Hhlch ties a s u b s t a n t i v e farmula to auudl if ier one, Is clearly more syntacticky,

L"

L

The auxiliary of an action also has its formula made dependent onthdt O f +,ne amrow iate action and the fact scored, but gheauxilliary formulas are not listed as dependent formulas eliher, 97s~are picked UP by fXfEJ3 and examined to determine the tense of heaction,TheY are then forgotten and an elwnent indicating the tense i sCONW onto the action: formula,In its initial state the wste~ wi I Irecognise only four tenses of complex actions,

PXS: ooes hide/is hiding/did hide/are niding/anl hidinnpFE: w a s hidfnglwere hiding/FAST : d i d him/had hiddenFUTL: will bide/wi)I be ‘hidiqg/shall hide/shall be hIdi-

24

In the cese of the negatfve of any of these tenses the word Vat" fsforgotten, and an atom NPRES, NIMPE, NPAST 0~ NFUTU att,athsd to ih?approptlate actlon formula Instead, At present the system does notdea I with passtves, though I Indlcrate later how they are dealt withwithln the template format,

Even when the ropresentatlon with the densest dependen0y he8 beenfound, there may still be more than one repressntatlon with thaiscore for a given fragment, So, In the ease of "The man lost hisI @Q" there may well be two rePrimentat\ons of this sentonoe with khesame dependency score, one corresponding to each of two differentsenses of *'legQ---one as a part of a body, and one as an inanimateihlng that supports some other thing (as In "plan0 IegV, There lsa further roWtIne In EXTEND, oalled lnto play in suoh terse+ tha;atteflpts to rstabllsh addltlonaltVsemant10 overlap" of oont8nt bothbetween the actor and object formulas of the template, and betweeneaoh of the three maln formula8 of the template and Its auallflers,If any can be found, the addltlonal dependsnotes ape u ad io choi)geamong reprossntatlons that heve rchllrwd the same soore Y nthe EXTENDroutines de8crlbed earller,So, In the present easer the formule forVeg of a parson )* would be expeoted to oontaln the aubfopmula (MANPAR?), whereas the formula for Van0 I eg'* would not, and thlaconneotlvlty wtth the lnltlal formula of the template, whose head was‘MAN? would 8Ufflce for one rePtrsentation to be ohoeren Inyeferegeeto the other8 agaln on the wlnolPl8'of PreferrIng the most oonnectedrepresentation,

Not ANY Oo-otourrence of olement$ would rrufftoe for th!s pirpoll,@ J 01oourss~ and an lmpOr&ni open aucrstlon in any 8Ystrm Ilk8 the presentone 18 what cbmblnations of element8 ate adeau@ir for ihepreferential selection bf formulas In suoh oaZ)@s,An example of aoomblnatlon of markers that 1s orrtaln to be slgnlfloani for iharesolUtl0n of rmbfgul‘y would bo (FLOW STUFF), a c0nyentloneloomblnatlon used to t ndlaatc, the ooncopt of fluids,So thrn ) fn-reaolvlng the possible amblgulty Of fntorbtetatlon of the sentenoe"The t a p Is drIppIng" we would sxgect to f\nd that obmbinatlon ofmarker8 present In the APPROPRIATE form

Ylasr for Vapf' and "dr ppln )V

and SO ta 8oloct the correct J PInterpreta on with their aid--y- n th sway-we would be able to dl80ard the “meat fatv sonaq of ndiipplng",*The. third and last pass of the text rp~lhs the TIE routlnerr, whhhestebllsh dependencies between the representattons of differentfrauentw, Eaoh text fragmrnt has boon tied by the rbutlnesde8crl bed so far to one 0r more full t8mplaiisr each oonsistlng ofthree maln formulas to each of whloh a llat of dependent fqrmulas ma?be tl@dt The Inter1 lngurl rrprosontrtion oonslst;;u:or eaoh textfragment, of ONE full template together with UP to addltlonalJtems of lnformatton called MY, Mark, Ca8e and Phaserespectlvely,The Int8rllngUal represrntatlon also 0onZalns {heEnglish name of the fragment ltsrlf,

25

The KeY is simply the first word of the fragment, lf it occurs on thelist of key words;or, In the cases o f Yhattq end v?whlchrt a key USE ofthe hard,

The Mark for a given key 1s the text word to which the key word tjesthe &hole fragment of whfch It Is the key,So, fn (He came hQmcl)(frOmthe War), the mark of the second fragment Is Varnefl and ihe secondfragpent Is tied In a relation of dependence to that mark by the key"frOt7", Every key has a correspondfng mark , found by TIC, unless(a)the key Is q*andqq or "but" or (b) the fragment fntroduced by "hekey Is itself a complete sentanoe, not dependent on anyihlng outs de4itself,The nOtiOn w;II become clearer from eXaminfnQ the sxameleparagraph sat out below,

From the Point Of view of the Present system of analysts, ihe Case oia fragment, if anyl genera!lY e%Dresses the role of that fragment Inrela4+n to its key and mark:lt speclfles the SORT of dependence, thefragrrent has upon its mark,

There ls one fmpgrtant case, OBJECT, whose assfsnment to a cast) doesnot oepend on the Presence of a keY,So, h the sentence (1 went Ifherto ieave) the latte r fragment would be assIgned the case OBJECT endwoulc' be tied to the actlon **want" as the mark of that fragment, eventhough there Is no key present'

But in general Case markers are attached to fragments on $he basis ofthe key a n d the mark It may be that no ease Is fInally assigned to afragwent, though It wtll b e If a fragment is lntroducsd by apregosltlon, The cases are, In a sense 4 a cross olasslfiuailon ofPreposWons p whose correct renderIn Into4 sey# French is so vitalfor adequate translatfon,for e%amDlq the Ensllah PrePbs!t!onOUTGF (8WeeZ8d into a slnQ(e Item bY the FRAGMENT routIn@) cap berendered Into French in at least seven ways,

The Provlslonal workInS list of cases and the English p)ePosltionsthat can Introduce them Is as follows;RECEIVER:to, from, forINSTRUMENTAL!wlth, byOIREC:TION:to, from, towards, outof, forPOSSESSION:wlthLOCAtIONbPace and ttme):at, by, near, after, tn, during, befopCGNTAINMENT:lnSGURCE:outof, fromGOALrtotat

The case analysis routines in TIE work by conslderlng t h e aboveclasslficatlon of Preposltlons In reverse 4 4s It were,So, fn (Hestru& the boY)(With a stlck)iTIE locates the With" and flndg In $hestere0tYDes for "w t t h " (see below for a descrfDtlon of sterotyoss)that%! th" ten Introduce either a POSSESSIVE or IVSTRUMENTALfragrrent, It reads there that If, for examPIe, an INSTRUbfENTAL cese1s in auestlon lt Will expect a PrecedfnQ actlon whose head 1s DO,

. . .

2 6

CAUSE or FORCEtand will also exDect a substaniive In the fragment Iilntroduaos Who80 head 1s THING, I n the paso mentionsd lt findsthese conditions sailsfled, sfnoe ths head of the appraprlatr iormulafor "at i ok" Is THING, and so ties the second fragment to ihs mark"h l t " and asslgns the INSTRUMENT&L case to the seoond fragment as adescrlDt?on of that tie,

I n any o t h e r sltuat\on, where t h e s e ctltatbt at@ n o t sailsi!sd, ;hefragment introduced by W*wlth~~ Is tied to the Immediately Dreosdinsrsubstantlve t and the case POSSESSIVE Is assl9ned ta the tlet as Inthe (He i9trut.k the boy)(wlth I on9 halA where the head of iheapDroprlrte formula for "halrfq is STUFFAn one special oiass ofcases5r the POSSESSIVE case Is asslgned even though a THINGsubstantive Is found fn the *'obJeot Dosltlon" of the seoond tsmPlat0followln9 on a D O , CAUSE or FORCE aotlon In a_ Drsoodfnatemplatr,Thoss are the cases where the obJeot Is a

0art of ihs

substantive prsvlously mentlonod,For~ even though a les s a T H I N G w ewould want to assl$n a POSSESSIVE oasl) to the second template of {hapair (HO hit the boy)(wtth the wooden le91,How thls TIlj IS obtalredalgorlthmlually Is dlsoussad in detail in the flnal sectibn of thepaper after the descrlpilon of STEREOTYPES,

This Drooedure oan be thought of a8 an amblgulty resoluiibn of GheDreposltlons, whloh was not been dealt with at at all by iha PICKUProutlnrs slnco Preposlilons are Insorted Into the formula sir]ngs aga sln9le formula and are never oonsldsred to b0 ambiguous a

fthat

staQ% the TIE routfnas also resolve ;$er 8smantlc ambipu ty notdealt with by the PICKUP routines, SOI l xamDle, If our last0XalVDb had b@en (He struck thr boyl(wlth a bar) we would haveexpected there to be at least two formula8 for "bar" stlll In play;correrpondln9 t o the head8 THING and POINT--iha lattercorreSDOndln9 to the pia sense Of qtbarq@,Honoe there would still betwo full templates matohlng onto ths latter fraomsnt at Ehlr stageand both oonsldered by TIE, whloh would thus Prefer the templateaontalnlng the sense of "barvv coded with the head THING, slnoe onlyln that case could a dependenoy tie ba made (to c~hlt~~ ln anotherfrigment In this case) on the basfs of InformatIon extraot@d from theformulas~ and In dofng so the ambloultY of “bar” would be resolvad,

Pha6e notatlon Is merely a oode to lndlorte ln a VerY general way toChe subueauent generatlon routlnos whore in the "progress of ihewhole sentenoefc one is at a given fragment A Phase number 1s aitrchedto eaoh fragment o n the follow~na bask by TIE whsrs ihs stagereferred to aDDlIes at the 6ECINNING Of the Cragmen; to whioh ihenumber attaches,

BemaIn subJeot not yet reachedl*subJect reached but not maln verb2*maln verb reached but not complement or obJeot3*ooRplement or obJect teaohed or not @YDWt@d

27L

cL

L

c

Anaphorlc lnformatlon of a fairly straightforward sort la DUi Intothe full template itself, So0 for examDIe, as TIE Dassas through anInput f;;fdit seek Dronoun formula8 and t@Di oet h e n the

sf;ylellmlnate alltemplate wlth the aDDroDrlat0 0sutptant ve

QOrtTUla---a the substantive to which the Pronoun refers ---trYIn asdo@* so to takeIt account of' a wlde range of exceptIons suoh as

fmsrersOna( us07

of pronouns t h a t it would be lnaocvopriate toreplaC8, as n yt seems that,* d'. Those uses can almost always bedetezted by their occurrence In company wkh a small and restrfotedclass of actions,

2,2)The Interllnsual Representation

L-

c-

What follows IS a shorthand vetston o f t h e InterllnQualrepresentation for a ;DaragraDh, desIgned to Illustrate the four formsof Information for a paragraph--key, mark, cass and phase--described above,The schema below 9lves only the bare temp!,aie formof the semantic informatfon attached to eachsemantic formulas and theft pendant Ilsts of formulas

fragmsnt----ihethat make uD

the full template structure are all omlttad,The French g/van IS onlyillustrative, and no jndlcatlon Is gtven at this point aa io how iii s produoed,

L-(LATER CM)

. (PLUS TARD VC)c ~nll:nll:nil:0tNo Template3

(DURING THE WAR Ch)L (PEYDANT LA GUERRF VG 1

CDURINCtGAVEUP:locatlonj0:DfHIS PBE ACT3c

(HITLER GAVE UP THE EVENING SHOWINGS CM)L(HITLER RENONCA AUX REPRESENTATIONS OU SOIR VG)

Jnl Itnll:ni J:ldtMAN DROP ACTJ.

(SAYING)L (DISANT)

CnllrHITLER tniI:3:DTHIS DO DTHISJL

(THAT HE WANTED)(QU'IL VOULAIT)

c-

L

tTHATISAYING:object: 3:MAN WANT DTHISJ

(TO RENOUNCE HIS FAVORITE ENTERTAINMENT)(REkONCLR A SA DISTRACTION fAVORITE)CTO:WANT:obJect:3:OTHIS DROP ACT3

(OUTOF SYMPATHY)(PAR SYMPATHIE: >COUTOF 1RENOUNCEI source v-1):-'DTHIS PDO SIGN3

28

(FOR THE PRIVATIONS OF THE SOLDIERS PD)(POUR LES PRIVATIONS DES SOLOATS PT )CFOR~SYMPATHY~recIpfentl31DtHfS PBE ACT]

(INSTEAD RECOROS WERE PLAYED PO)(A LA PLACE ON PASSA DES QISQUES RT)CINSfEADtnlltnll:0:~AN USE THINGl(comment;tem~laCe rcthe)

(BUT)(MAIS)EBUT:ntltnll:O!No Template1

( ALTHOUGH THE RECORO COLLECTION WAS EXCELLENT CM)t BIEN WE LA COLLECTION DE DISQUES FUT EXCELLENTE VWfALTHOUGH:PREFERREO:ntl:OrGRAIN BE KIND3

(HITLER ALWAYS PREFERRED THE SAME MUSIC PO)(HITLER PREFERAIT TOUJOURS LA MEME MUSIQUE PT)Cnjl!niltnll:OtMAN WANT GRAIN3

(NEITHER BAROQUE)(NI LA MUSIQUE BAROQUE 1CNEITHERtMUSIC~auallfle~~0:DTHIS DBE KIND)

'(NOR CLASSICAL MUSIC CM)(NI CLASSIQUE VG)fNOR:INTERESlEOln~l:0;GRAIN OBE DTHIS].

(NEITHER CHAMBER MUSIC)(NI LA MUSIQUE DE CHAMBRE)CNEITHER:INTERESfEO~nll:OrCRAIN DBE DTHISJ

(NOR SYMPHONIES CM)(NI LES SYMPHONIES VG)-~NOR~INTERESTEO~ntI:0~GRAIN OBE DTHISJ

(INTERESTED HIM PO)(NE L'INTERESSAIENT PT)Cnl.ltnllrnll~l~DTHIS CHANGE MANJ

(BEFORELONG THE ORDER OF THE RECORDS BECAME VIRTWJ,.Y FIXED pD)(BIENTOT L'ORDRE DES DISQUES DEVINT VIRTUELLEMENT FIXE PT)CBEFORELONGtnfIvdl:i?!:GRAIN BE KIN03

(FIRST HE WANTED A FEW BRAVURA SELECTIONS)(O'ABORD IL VOULAIT QUELQUES SELECTIONS DE BRAVOURE)Cn~l~nll:nll~B~MAN WANT PART1

(FROM WAGNERIAN OPERAS CM)(D'OPERAS WAGNERIENS VC)CFROM:SELECTIONS: source:?rOTHIS PO0 GRAIN2

c

c

Ir

t-

11L

L

!

29

(TO BE FOLLOWED PROMPTLY)(QUI DEVAIENT ETRE SlJIVIES RAPIDEMENT )CTO:DPERAS!nll~3:MAN 00 OTHISg(oommentishlft t o aotfve templateagain may give a different but no-t lncorrrct translatfon)

(WITH OPERETTAS PD)(PAR DES OPERETTAS PT)C~ITH:FOLL~WED:n~~:~:DT~IS PBE GRAIN3

(THAT REMAINED THE PATTERN PO)(CELA OEVINT LA REGLE PT)[nil ;nfI:nilt0:THAT BE GRAIN3(comm@nt:no mark beoause @$hat' tiesto a whole sentence,)

(HITLER MADE A POINT OF TRYING)(HITLER SE FAISAIT UNE REGLE D'ESSAYER)~nll-Wl:nii~~~MAN 00 ~THISWornrn~nt Isome ldlom reoogn;tlon ess,ntlalt0 cope 4th thl,)-.(TO GUESS THE NAMES OF THE SOPRANOS)(DE DEVINER LES NCMS DES SOPRANOS)TfO:TRYING:obJect:2:DTHIS DO SIGN!

(AND WAS PLEASED)(ET ETAIT CONTENT)CAND:HITLER:nllt3:OTHIS BE KIND3

WHEN HE GUESSED RIGHT CM)(WAND Il. OEVINAIT JUSTE VC)CWHEhtPLEASEO:location;3:MAN DO OTHIS3

(AS HE FREQUENTLY DID PD)(COhME I& LE FAISAIT FREQUEMMENT PT)CAS:GUESSEOtmanner:3:MAN DO OTHfs

It 1s assumed that those fragments that have no template attached totherr ---such as (LATER)---word-fopword means,

can be translated adequately by puralyWV8 It not for

r8adbW it, W8the ditflcultY involved in

could IaY out the above text so as to dfgplay the;1~~~~d”““:~oS,1”P~~~~dbYa~~8 ~~9h3tmeW of cases and marks at the Word

partidular words,SoOf deoendenafes of whole fragments on

P for emwle the Watton of Just the first Gwofragments could be set out as followst

IOTHISJ - during - war *the4* (location)4l

hltler - gave+up - showtngsc thet

iventng

30

The interlingual representation described 4 as the result of Sheanalss1s of Engllsh text, a n d fllustrated a b o v e h b a r e templaie form

1s t h e I n t e r m e d i a t e f o r m h a n d e d , a s It worekalysls prodrams to the Frenoh generation ones,'

from the Enoll8h

Howeverr this lntsrmediate stage I"SP as It must be, an arb!irarY heIn the English-French orooesslng that It la W~tul to examine at {hesurface level here for e%oosltorY purooses and not on/Y fn ihe oodedfOrt5, There Is often a misunderstanding of the nature of antnterllngua, In that 1t Is supposed that an lntermedlats stage ltkethe present Interlingual representatlon(IR for short) must aontafn"all Possible semantic Informatlontq 1n some exolleft form ff the IR1s to be adequate for any purpose,

But the quoted words are not8 and cannot bet wel I defjned w?threspect to any codjng scheme whatsoever, What ls the case 1s thatthe IR must oontaln sufficlrnt lnformatlon so as to admli; oi formalmanlPUlatl0ns u p o n itself adequate for Produolng translatlons Inother natural or-formal Ianguages,But that Is suite another maker ofoourseo

The fal lacy 1nvOlved Is analogous to thai; committed by Ghecomputstionally 1lllterrte who say that "You cant get more out of acomputer t h a n Y o u Put In, can You?*'------- whloh Is fake If it istaken to exclude comPutatlon upon what You aut In, (A mbretradlt10nai parallel 1s the Socratiu argument about Wheiher or noithe pramlses of an argument vvreallY~’ oontaln all possible conclualbnsfr0R themselves already, In that to know the promIsses Is already toknob the conclusions),

Anaio9OuslY, the IR for translation need nolt; contain any aartloularEXpLICIT information about a text,The real restrlctlon ‘fa that increatfng the IR no fnformatlon should have been thrown awaY that willlater turn OUt to be fmpOrtant~ So4 if 000 makes the QuPerflcfal butcorrect genetalisatlon that one Of the d1ff1cUit1ss of EnglishmFrsnohMT Is the need to EXTEND and make exollolt in the Frenoh ~hbw$ thatare not so 1, the English, then It Is no answer to say there h noorob lem slnoe, whateve,r those thlngs arel the IRr If adequate, mustoontaln them anyway, It Is then argued that If there Is a d;oblrm iiIs -a general one about dsrlvlng the IR from English and has nothjngat all to do with Frsnth,

But this, as I have polntsd out, need not be true of a n ypar$loularIR, since any IR must be an arbttrary Wt off stage In gojng from onelanguage to anothsrla slioe taken at a partlouiar dolnt forexarinat1on, as it were,

Consider the sentence Vhe house I live In Is tollapsfng*~ whiohcontains no subJunct1on Yhat% though in French It MUST be exPressedexplicitly, as by “dans laauel W, There need not be anyrepresentation Of "thatft anywhere In the IR,All that Is neoessarY fsthe subordination of the second fragment to the mark *thousV Is

31L

cc

c

a n dcoded, generation procedures t h a t know In such casas ofsubordination an aporooriate subJunctIon must ocour ln the FrenchOUPM, It I S the need for such procedures that oonstltutes 6hesometimes awkward expansion of Enslish into French, but the need fortherr IN NO WAY aictates the explicit content of the IR,

G

L

Lc-

2,3)The Dictionary format,

The dictionary is e s s e n t i a l l y a list of Pairstof semantlc formulas,(each corresponding to one sense of an fngllsh word), and ofexplanations of that sense,By )( eXplanatiOnv 1 mean not simolY anEnglish word or Prase, such as was used tn earlbr versfonp of thbsystem of analysis to distinguish each sense from others, but what Ishall call a French STEREOTYPE,

L

c-

In earlier versions of this method of analysis Cl5 1 one sense of,say, the English word%oloriessv*might have appeared ln the dl&Onaryas:

(UWJHERE SPRU-CWSENSE SIGN))NOTHAVE)KIND)(COLORLESS AS NOT HAVING THE PROPERTY OF COLOR))

IL The first half of the oalr, the formula, 6xpr0sses t h e faci thai.bein0 colorless is a kind or sort whloh mean8 not having a spatfal

L(WHERE SPREAD) sensory proPettY (SENSE SIGNLThe second half of the

L i3atr Is a sense explanation In Engllgh that contains the name of thewora avd serves to distinguish that Particular sense of ~colorl0ss"frar other senses---such as one about human character--- for anyonereaaing the dictionary who was not familiar with the codfng sYsi0mL embodled in the semantic formulas,

1 But, the senses of the English words d~stingulshed by the d\&onaryL may equal IY well be explained and dlstingu~ished by means of thr I r

French equivalents, at least, In cases where ihe notion of" a Frenche equivalent to an English word" is an aPPro~rlaie one,So, for example,the French words *'rouge" a n d "sociallste" might be s a i d t o

L-distinguish two senses of the English word 'redv, and we ml ghi codethese two senses of "red in the diet iOnaryv by means of the sensepal&

LUMHERE SPREADIKINDURED (ROUGE)))(((WORLD CHANGE)WANT)~~AN)(RED(SOCIALISTE)II

The French words f@rouge" and tQsoclal istaft a r e enclosed In listcarentbeses because they need not have been, ag in this Gage, singleFrench words,They could be Frenoh words strings of any lengthtforexarrpie, the qualifier sense of r*huntingt~ a s It occurs In a “ahunting gun" is rendered In French a s "de dttssewr henoe we wouldexoect as the right hand member of one sense pair for "hunt ln~~~(HUltTING( OE CHASSE)),

32

Th i s simpiified notton of StereotYPe Is adecluate ?Oi iherePresentatlon of most QUaI ItierS and 8UbstantiVesdeh~ ,I ah&generaiise to the notion ofrepresentatl0n Of

a FULL ST E R E OT YPE adeQu@te for t h eprepositions and actbns, In whiah the;, maY bs

more than one list after the Engllqh word name ?n the ifght hcndmember o f the sense Palr,Moreover, they will be Ihtl, in whlohfunctions will occur as well aa the names of French Words,

But we should pause at this Point Just long enough to see what ihsnotions of sense PB!~ and StereotYpe are doing for US bl the sY8tem.Earlier on8 1 desertbed the structure of a full template ----assfgnedt o some natural language fragment-----as made UP of (rbrmuirs andlists of formulas,But these would more aocurately have been desorigedas sense pal rs, and ilsts o f sen8e lairs, That Is 20 S~YI sheanalysis routines in fact build into the tenwlate not Just theformulas but the wttOLE SENSE PAIRS, of whloh the formula8 are iheleft hand members, even though the crftera for lnoorpora$lng a 8ensapair Into the template apolled only to the formula Itself,

Hence the ful I --' ternPlate already oontalns tho French eaulvalents 04the Enollsh words lk t h e fragment,Moreover t h e ste;ebtypes foiactions and Preposttlons oontain not only French equ\valent8 butlmpllcit rules for assembling these equivalents SO au to generateFrench output, Thus the generation routtnes never need fo cknrult anEngll8h-Frenoh dkt]onary,Ail the ganrratlon Orogram re~ufra8, tnterm8 of Frenoh eQu\vaiOntS and assembly rules, 18 already die8ent Inthe full temolate,

Thus the full template may appear to be a oomplox and Oumbrou8 l;jemof informatlonr contalnlng a8 It doe8 not only a oonoepiual 8rmanZlorepresentation of English text8 but aiS Frenoh outpui fbrms a n dlmpl icit generatlon rule% But the avoldanos 03 rePea<edoonsultation of a large dlotlonarY of forms and rule8 ln LISP fofmaiIs no spall OomPensatlon,

ihe full stereotype thsnr may oontak not only Frenoh words but al8opredicates and funottons of InterlIngua! I tern8 whose valutrr rrfjalWaY French word strlnss, or a blank Item, Or NIL, The no$lon oflnterllngual Item hers Covers not only the intsrltntiual slrmrnts thatmakb UP the formulas, but also the names’of the G;;;; abbieviated toa standard four letter format #for sxamolerRECEl 01% PDSS,LOCA 8 CONT, SOUR, COAL, OBJE, Q U A L (300 ihe Ilst'of ba8e8 QtvenearlIe&

the general form of the stereotyoe Is a list of Dredkate8, toi lowedby a rrtring of French words and funotions ihat evaluate io FrenohWordC o r to NIL (?n whloh oaso the srterootype fills), Thefunctions maY also evaluate to blank 8YmbOl8 for reasbns to bedescribed,

33L

The predicates -----which oocur only in oreoositton stereotypes---k norfV!W~ refer to the case of the fragment containing the word, andc to its rark respectfve(y,lf both these predicates are satkfled the

program continues on through the stereotype to the Frenoh output,

L Let us consider the verb "-a d v ; s e " , rendered in its most

etstraightforward sense by the French wad "conseiiier,,, ft is Ilkslvto be foliowed 5y two different constructions as in the Englfshl i) Iadvise John to have patience ii) I advise patience

bL

Verb stereotypes contain no predicates, SO we might expect the mostusual sense pair for ,radvise,' to contain a formula followed by(ADvISE( CONSEILLER A (FN1 FOLK MAN))

( CONSEILLER (FN2 ACT STATE STUFF)))

LThe -role of the StereotyPes should by flow be becoming c l e a r Iingenerating from, ’in thls case, an action, the system looks down a

c-list of stereotypes tied to the sense of the action fn the fulltemplate, If &ny of the functions it now encounter8 evaluate t0 NILthen the whole stereotype containing the function fails and the nextis trled,If the functions evaluate to French words then they are

L generated along with the French words that ar>pear a8 their own names,. like Vonsel 1 ler",

L The details of the French generation Procedures are dkoussed insection 2,s below, but we can see here in a general way how thestereOtYPes for "advise" Produce correot translations of 8entences(1) and (ilL In the case of sentenoe (I) in the form of iwof ragrents (I advlse JohnHto have patience), the p;;gt;;; beglns togenerate from the stereotype for the formula I? the oositionin the first fragment's template,It moves rightwards as described and

L begins to generate "consei l ler a yfhen (FNl FOLK MAN) 1s evaluated,which Is a function that looks at the formula for the third, obJ@ct,

m position of the current template and returns ITS French stereotypeonly If Its head is MAN or FOLK -----that is to say if it is a human

fL-being that is being advised,The formula for "John,, satisfies this and9’ J 8 a n v is generated a f t e r Vonsell ler a,,, -------proper names aretranslated here for illustrative ourpo8es only-----and SO we obtainthe correct construction "$0 consei i le a Jean,,,

c-

L-

But haa we been examining sentence (ii) ,,I advlse wdenwP thisfirst stereotype for "advtse" would have failed sinoa (FN1 FOLY MAN)would not have produced a French word on being applled to ihO formulafor Vat/ ence", whose head is ACT,Hence the next stereotyoe wou I dhave been tried and found to aPpIY,

The stereotypes do more than simply avoid the axpiicit use of aconventional generative grarrmar (not that there is much precedent forus I n9 one of those) In a system that ha8 already eschewed the uee ofan analysis grammar,They* also direct We production of the Frenchtranslation by provfdfng complex context-sensitlva rules at the POtnt

34

reauited, and without any soaroh of amethod I S,

(atgo rule kivonioiYJhf8fn prtncfple, wwmlble to the oioduotton‘oi ~raronablu

oomp(ex lmollclt rephraglngs and expansions, as tn the dorlvrtlon bf"8 I lntelllmnt salt-11" f r o m t h e 8,aoond ftrsmont 01Hhowever lntellIgent)(can survive death),

(No mangtven the aidropriaih

stereotype for "hoWever@*,

Preposition stereotypes are more oompiqx In general than ihbsa io;actIons, but before IIlustratlng them I should mention a dolni ihatart808 In oonnextbn wlth stereotypes and their rela$ifbn to iheenufleratlon of the senses ofdescribed the dlOtlonarY so

Input Enollsh 1 wordr;As I h a v efar, many output ster@otydos may be

attached to one sense of an Enollsh Word8 that Is to say to a singlesemantic fbrmu(a, In the e%aWe sentences abbve, "adv SV 1s takenfas being used In the same sense fn the two sentenoes,affferent constructjon follow the word in the tWo cases,

even though

So ihe notton of StereOtYPe In no WaY oorresponda to thai bf Wbrdsense, Indeed, >hs notion of Word-sense Is an extremely unolerr bneand rSSi8tant t0 any formal ana&sls,Wlthout In any way olafmIng thaeihe senses of a Word can be oompMe(y enumeraied, It Is nonegheles8olear to oommonsen8e that In "1 have a bar In my new h&se" and "wehave a bar against foreIgnera here".iwo different senses fn terms of

the word flbar'f Is being used In

even though It la nbt possible to~~oommtual fpparrtton of pbr+yts*f,

terms of naive denoiaiionreWloat0 ghat last ,oonorrot in

Or formal 8Peolflcatlon of tontei)ls,

In the case of preposltfbns I take them as havjng only a slngls ssnsseach, even though that sen8e may olvo rke to a greai number oistereotypes,Let us conskl@r, bY way of e%ampJe, woutofw(oonsldsred asa slng(e word) in the three sentenoes;

1) (It was made)(outof wood)ii) (He kIlled hlmI(outof hatred)-tif)(I live HOutof town)

It seems to me unhelpful to say that here are three senses oi ~~outof~~even though Its ocourrenoe In these examp(e8 reclutres banslntlonIntO French by v de,', tqparO and "en dehors deft respeo~lvrly&nd othereontexts would requtre wparml" or "dans?

Given the oonvsntlon for stsrsotypss dssorlbed sarllsr ibr acttons ,let us set down stereotYpe that would enable us to deal w?th thesecases I

Sl) (WRCASE SOUR)(PRMARK @DO) DE (FNl STUFF 7HING))SW (WRCASE SOUR)(PRMARK *DO) PAR (FN2 PEEL))Sill) ((PRcASE L0c~> EN DEHORS DE (FNl POINT SPREAD))

Where *QO lndlcates a wide class of action formulas; any In faotwhose heads are not PDO, QBE or BE,

L. 35

LL

CL

In the case of the sentence fragments (It was made ) folJi;of wood),when the program enters the second fra$ment lt knows from the whileinterllmual tepresentatlon described earlier that the case of thaifraarrent Is SOURCE and Its mark 1s @gmade*O, The mark word has DO asits h e a d , and so the case and mark predloates PRCASE and PRMARK Inthe ffrst stereotype are both satlsfled,Thus (( de" 1s tsnattvelygenerated from the first stereotype and FNl 1s wpl led, because ofits detlnltlon, to the object formula In this template I that 1s tosay I the one for %oodYThe arguments of FNl are STUFF and THING andthe function finds STUFF as the head of the formula for "wood" in thefull template' iS satisfied and so generates " b o l s " from thestereotype for "Wood"'

c

L

C-

i

In the case of the second fragment of (He klllad hlm)toutof hatred)the two predicates of the first stereotype for "outof" would again besatisfied, but (FNl THbjG STUFF) would fall with the formula for"hatred" whose head is STATEJhe next stereotype Gil) would betried; the same two predfGates would be satlsflsd, and now (FN2 FEEL)would be wpI~,ed t o (NOTPLEASE(FEEl. STATE)) the formula for“hatred”,But FN2 by Its definltlon examines not formula heads # butrather seeks for the containment of one of Its arguments wfthln ihefortWla,Here It finds FEEL wlthln the formula and so generates theFrench word stereotyoe for "hatred"

ISimilar conslderatlons apply to the thlrd example sentence fnvolvlng

c the LOCATIOtv casetthough In that case there would be no need to workthrough the two SOURCE: stereotypes already dlscussed since, when acase is assbmed to 8 fragment during analysis, only thosestereotYpes are left in the lnterltngual representation thatLcorrespond to the assigned case,

L The description of the asslgnmeht of case to a fragment was deferredL fror the earlier discuss/on of TIE routfnes, slnce 1-t requires Use of

the stereotypes at the analysis stage,In the CBS8 of fragments Wfth a- key, TIL routines search the stereotypes for the key until they fbd

LL

one that matches the fragment and Its mark except In respect ofcase,So, in the sentence (I live)(outof town) the analysis routinesasslgn LOCA T I O N t o the second fragment in the first place beaausethey locate In the third steretype for "outof*~ a formula for thei. object of the Preposition whose head is POINT,

L‘2,4) The generation of French

Much o f tne heart of the French generation has been described lncoutllne in the last section' since It Is impossible to describe thedict!OnarY and its stereotYPes UsefullY wlthout describing the

e generative role that the stereotypes play,

. .

36

To complete this brief sketch all that It 18 appropriate io add issome description of the way in which generations from the stereotypeof a key and of the mark for the same fragment Interlook-;- the markoeing In a d i f f e r e n t fragment---as oontrol flows backwards andforkads between the stersotypss of. different words in asarM of Qsatlsfaotory French OutPut, There is not spaoe availab'is.hrrs fo;&;;;f;tion of the bottom level of t h e Qeneratlon program--the

and number routines-- which in even the simplest oases needacce88 to mark information, as in iooating the gender of/heureux~~,in(John seems Hto be happy) translated as "Jean ssmble etre heuteux%

Again, muoh of the detai ied content of the Qeneration 18 io be foundin the functions evaluating to Frenoh words that I have alb&arllyn a m e d FNl ,,,.etc,Some of these seek detail down to Qend8r markers.For oxample, one would expect to Qet the oorrect translations "$0voyaQeal8 en France” but “. , #au Canada W with the aid of ?un&ions vsay, fNF a n d FM that seek not only speolf lo formuia heads butgenders a8 we0,s0,would exPeot to f=,lnd (g ven that formulas for land areas havet

mane the ster~otYtw8 for the English 'rinVt weSPREAD

a3 their heads):SPREAD)),

I ,,,,,,,,A (FNM SPREAD)) and m ,,,,,EN (FNF.

It Is not expected that there ~lil more than twenty oh so of theseinner stereotype functions in all, Though It should be noticed aithis Point that there Is no level of generatlon that does not reauireaulte oomplioated semantic information processing,1 have In mind herewhat one mloht call the bottom Ievsl of QeneratlonAhe addition andoompresslon at artloles e An Ml program has to get WJe bois du vin"for '@I drink wirW but to "J'alme LE vln" for "1 iika Wine~~~Now thereis no anaiog for this distlnotion In English and nOthing about {hemeanlrlgs of Vlksr@ and "drinkk* that accounts for the difforrnoe hthe Frenoh in a way IntuitfvslY aoceptable to ihe English sdsaker',Airepresent we are expsotlng to generate the difference by msans ofstereotypes that seek the notion USE In the semantio oodings --whiohwill be looated In @qdrlnkVf but not in "llke?and to use this togenerate the ))de" Where appropriate,

The overall oontrol function of ths generation expeots five diffsrsnttypes of template names to OOcUrt

W&IS *DO *ANY where *THIS is any substantlve head(nbt DTHIS)*DO Is any real action head (not BE, p&p 1)BE)

and *ANY is any of *DO or KIN0 or DtHIs,

Yith thls type of ternpints the numb or, person and Qe nder of the vrrbare deduced from the Frenoh ster eoty PO for the subJscr t pa rL

la) type *THIS BE KIND 1s treated wfth tYPe 1,

2)DTHIS *DO *ANY

G

c-

b-

i

Lt

L

37

These templates arise when a subJect has Peen sPllt from its adonby fragFentattan,ThR mark of the fragment is then the subject,Or, ihet empla te may r e p r e s e n t a n obJect action phrase, suoh as a ahdieintinitlve with an fmplicit subJeo-t to be determlned from the mark,

3)+THIS DBE DTHIS

Templates of this type rwresent the subject, split off from itsactlon represented by type 2 template above, The translation ISslmp IY generated from t h e stereotYpe o f t h e 8ubJect formula, slnoethe rest Is dummies, though there may arIse cases of the form DPWISDEE KIND where generation Is only possible from a qualiffei as ln shesecond fragment of (I like tail CM)(biond CM)tand blue-eyed Germans),

4)DTHIS PO0 *HEAL

Templates of this type represent phrases and &hettanslatlon is generated as described from the key stereotype1 afterwhich the transla'tion for the template object Is added (*REAL denotesany head In *THIS or 1s KIND),

The wneral strategy for the flnal stages of the MT program is t0generate French Word strings dlrectiy from the template struotureasslQn8o to a fragment of English text, The first move fs io f!ndout hhtch Of th 8 five maJoT types of template dlSt~nQu~8hSd above ISthe on@ attached to the fragment under examlnatlon,

so then, for a fragment as simt3l~ as "John already Owns 8 big iedoar", the program would notlce that th8 fragment has no mark or kru,hence, by d e f a u l t , the generatlon Is to proceed from a stereotypewhich ls a function of the general type of the template at);achlng tothe frasment,The bare name of the template for this one fragmentsentence la MAN HAVE THING and lnsPeOtiOn of the types abova Wfllshob thls to be a member of tYPe (lh whose general form la *TWIS *DO+AN'I,The stereotype ts a function---let Us 9aY FTEMP--*I of thatt e m p l a t e type and to conform withstereotypes described kariler,

the wweral format forthla can be thought of as being one of

the stereOtYPes for the "null word", since we have no mark or keyword -to start from here,

In this case the generatlon of French Is slmpliclty itsalf:{hefunction FTEMP evaluates to a French word string whose order 18 thatof the stereotypes of the Ensllsh words of the fragment,This order 1sdirected by the Presence of the first type of template oomprisin9 anelementary sequence subJect-action-obJect,This Is done racurg/vely sbthat 8 alo+ with the French words generated for those EnQIISh Woi$whose formulas constitute the bare temalate(l,e, *'John", tr0wnffVarV are generated those whose formulas are merely dependent on ihemain formulas of the template--- In this case the fo;mulas foi~'aIreadY", trb ig?f’ and "red*', d

38

If complex Stereotypes are located while Oenetatlng for any of $Wword8 of the f ragment--0-"oomPleX" 8lmplY means full 8iieraocYdeswhich have oonstltuents that are funotfons as well a8 Frenohwords -----then generatIon f r o m thare newlyfmmedfately takes Drecedsnoo

?oynd p~ertoiuiqoV!r further gwwatlon from the last

stereotype at the level above,

In the Present cage wown~" create8 no problems slnoe It Is aoompletely regular Frenoh Verb@ and so Iis stevoiydes oontafnnothing but Frenoh wOrd8rh sWWah If 18 bnlY Irregular F t e n o hVerb8 that aontaln comDle~ltY In thalr stereotypes 80 88 to $fotCtethe form of what follows them 1n a sentenoe,(It crhould b8 undW8toOdthat I am u8lng "frregular " bore to meean Irr@gular Wfth r@SdbOt tothl8 sY8tem of classlftcatton ----my uaa90 18 not jgirndrd tboorrosoond to the standard opposltlon of vVegular~t to ~~!rregulaP fnFrench wammarsL

NOW suppose we oonsfdsr the two fragment 8entenoe 9 order Jbhn tbIeavVtThe fragments will be prearented to the gonerott& program Ifnihe f o r m dssort.bed earlier

i4th Keyr Markr Ca8e and Phase

jnformatlon attached to saoh ragmentl

(I order John) nlltnlltnllt0(to leave) tOtorder:OeJE:2

Also attaohad to the fragments will be full template8 whoso baretemplate naes tn this ca8e will be MAN TELL MAN and DTHIS MOVE DTHISie8peottvalyc

The generatIon prog+n enters the flr8t fragment which ha8 nk mark oikOYI80 it 8tarts to generat& a8 before, from a 8terrotyde for {henull word whloh agaln Is one for the flrrrt template tYie;Thfl getsthe 8ubJect right ; " J 0" from the stereotype for VQ 1a;t)ei to bemodif bd to 11 J I w by the conaord routIne,It then enieis ihsbtereotYpe8 for the actlontthe flr8t being( ORDONNER A (FNl MAN FOLK)) The head of the formu a forMAN, and FNl here 18 an arbkrary name for a funot on thlt looks intoI

"JohrV !,8

ihe formula for the obJect Plaoe of a template and, If gho h8ad ofthat formula 1s any of the funotfon'rr argumenta, it ;e$ui;nr ihe3teceotype value of that formula&! this chase t h e funbib FNl fssait8f Ied by "John", so bY deffnltion that 8terrotYpe tbr t$rder" issatisfied, and the program generato from It the seguenoe rlordknne; aJean"1 givlno the correct sequenoe !cJoS ordonners a JearV=-- where !§Indicates the need for further minor procrsslna bY -the aonobrdroutlne,The stereotype has now baen r%hau8ted-l-~~not~ino in 'ttremains unevaluated or ungeneratod---~ImllarlY the fraQm)rnt isexhau8ted 8lnoe n0 words remain whose stereotypes have not beengenerated, ofther dfreotlY or da the steieotype for som8 bihef woid,and so the program passe8 on to the 8eoond fragment,

39k

c *c

C

c

The tcvogram enters the second fragment and finds that ii has a mark,namely "ordcsr",It then consults the 3t3r3otype in h;;dwI;r "order"~)nfra9Fent W to see if it was exhausted or not, # and 3a theprogram turns to the stereotypes for "to", the key of (I\) Amongthose whose first predloate h5.3 the argument OBJE w‘ill be t h estereotype((PRCASE OBJEHPRMARK FORCE TELL) DE (FNINF *DO))

If we remember that the head of the current formula for "o~dert~ # Chemark of fragment ( fl), 13 FORCE' and that PRMARK seek3 and comparesfts arguments with the head of the mark formula, then the predltatesare seen to be satlsfted and the program generates "de" after seeing

c-that FNINF is satfsfied, since an aotion formula for "leave" follows,whose head MOVE is in the class 420,

FNINF on evaluath finds, where necessary' the implicit subJaot ofthe lnfinltlve,that fs unneoessary here,exam!es only sllghily more camolex,

but would be eesen$ial fnsuch a 3 "Marie reqretta d e

c-s'etr3 rejouie troP tot cr,Flnall~ FNINF itself evaluates to the Frenohstereotype selected for "leave",This might Itself give rTse to tomore gea,chln9 If the use of I* leave” dlctapd It, own $eQpents a, In"I order John to leave by the first trafn", Here however the

i evaluation terminates immediately to VartW since the sentenoe.stops,The program makes no attemot now to generate for “leave y*

L again' singe It realtses it has already entered its 3terr)otype listvia the v t 0 " SteteotYp@, Thulr the correct French string "Jo8ordonneg; a Jean de partIP has been generated,

The last example w33 little more than a more detalled rtbde8CrlPtlOnof the prooesses described In the dlotionary 8ectfan (2,3) inconnexlon Wth the example "1 advise John to have Pat ienoe", HoWeVer,now that we have dealt fully with a fairly standard case and ghownthe reourslve use of stereotypes In the generatfon of Frenoh on afragment-by-fragment basis, we oan dlscu33 a final pair of examples

e in which a more powerful stereotype, as lt wero# oan diotate and takeover the generation of other fragments,

L If be were to consider In detail the generation of Frenoh for the Iwof ragrrant sentence (1 throw the ball)(outof the window), we shouldfitjd the proce3s almost identioal to that used In the last e~ample,Inthis case, too t the mafn stereotype used to generate the Franoh forthe f I rst fragment Is that of t h e action---vpthrowtt In thiscase - - a n d the StereotYPe for "throw" is exhausted by the fl;sifragrrent, so that nothing In that stereotype cau3es the p r o g r a m t oIrlspect the second fragment,

Now consider, in the same format, (I drink wlne)(outof a gla3eLFollowing the same procedures a3 b e f o r e we 8hall find auraelvesprocessing the stereotype f o r "drink*' whloh reads ( BOIRE (FNl (FLOWSTUFF)) (FWl SOUR PDO THING)+ DANS (FNXZ THING)) where"?" Indiea~esa halt-point, The Drogrjrm beglne to generate tentatively, evaluatjngthe functions left to right and being prepared to canoe1 the whole

4 0

stet@OtyPe If any one of them fails,FNl is applied to ths foimula to;@*w f n e ') and s~ecifles the lnolusion in it8 formula, not of bns of iwo Lelements;STUFF),

but of the whole oonventfonai subformula for ,lfoujds (FLOWThis it finds, is satisfied, and so ovaivates to "vintf, ta

be rrodlfied by concord to "du vln*t,

The Program now encounters FNXl, a function which by deflnftionapplies to the full template for some FOLLOWING fragment,At thispoint the pto(lram evaluates FNXl which &urns 'a blank symbbl if andonly if It finds a fol/owlng(though notfollowins

neoessatiiY fmmedlately1 fragment with a SOURce coed and a template, the last Gwb

elements of whose bate name are PO0 THING, Le,type fragment with a physloai

It 18 a dirP0sCtlOnobject as the objeci: of

prepo8Itlon,ThIs siiuatton would n o t obtah if the sentence978

WlPO ‘1 Idrink the wine outof Politeness? If FNXl la satisfied, as h tblpeaser i t c a u s e s t h e g e n e r a t i o n f r o m t h i s StereotYPe tq bait after.geneyating a blank symbol, Haltlns in an evaluation 1s t&be takenas quite different from both exhausting (all functionsFrench word str.1 ngs

ova I uaied t6OP a blank) and faIlina fat least one funotlon

evaluates to NIL),

The main oontroi program now pagses to the next fragmen& In thfscase tWJtof a 9 lass”namely "dr I nk5

,It asks first if It has a mark8 whigh It hrsand looks at the stereotype ‘fn hand for the mark t6

see If is exhausted, whtoh it is nobiherefore continue8 to

merely haited@Thc, drogfamgenerate from the same sCere0tYP.r

“dr i n k ” , producing trdu vin", then vfdana~t ( to I lowed by ihqi0;

evaluateof FNX2, name ! y "verre", thus oivine the oorreot ttanslat!on "38boisS du vin dans un verre?

The important Point here 1s that the stereotypes for the ksy io iheseoond fragment "outot", arettanslatlons for

N E V E R CQNSULTED a t alI,Theal&e words of the seoond fragment wfll hav@ been

entered via a stereotYPe tar the prevloue fragment, ths one to;"drink",The advantage ot thle method will be oiear~becauee It wouldbe very dlffioult, obn@eptualIY anddescribed

wlthln t h e framewofk I h a v eI to obtain the tFanslatlon Of V@outoCff as “dansf’ in thfs

oonte%t from the stereotype tar Voutof% beoau8e that trans laif qn isspeaiflo to the occurence of c e r t a i n Frenoh Words, s u c hr a t h e r than to the applloalon of oertatn oonoepts,

aa,- Vtbolran,

stereotypes oan ooze wtth lingufstio Idiosyncrasy asIn thfs way the

oonceptual regularity,ItWI I d h

shou Id be noted, to& that sinoo $ncln t snot generated unt!l atter the halted stereotYpe restatts, zhere 1s norequirement that the two example fragments be oontiQuous,the method Ih a v e d e s c r i b e d oould cope J u s t a s w e l l with (I drink t h e whe)(I llkemost)(outof a sliver goblet),

The Point here ( a b o u t what worda we fJenerated ihiouah ;hostereotypes for what OTHER words) can Perhaps be made lIttIe oleaieiwith a diagram in which Ilnes oonnett tho Engil8h word throughwhose stereotype a generation is done to the word for wh?oh ouiput fs

41L,

yenerated,Ali generatlons conventionally start from 0 the null ward'Ic mentl Oned $r)OVO# it t&bY convention, the word for whfoh the f!veL bas i c stereotYPes are the stereotype' so then,the more

straightforward ewe (I threw the bali)(outof the window) would begenerated as follows!

- I throw-bail OutOf**+Wlndow

Artioles am omitted for simaltcltu,In this case the new fragmentstarting with "outof" returns again to 0 to begln generatIng aga'fn,In the trove complex case (1 drink w)ne)(outof a glass) the gensfstlonpattern would be as follows:

0

L

‘c

L

where the subJects andseparated bY lntervenlng

obJeotsolauses

of a sentence are obnsiderablv,these oenerat~on dfagiams can

beeowe consIderablY more oomplicated,

The generai rule wtth aotlon stereotypes then, is that the mbreirregular the action8 the mote InfOtmat\On goes into its stereotypeand the lessexarrpie,

Is needed In the strreotyoesf;;r Its seque!ts'Sor forthere is no need for a stereotype "outofr to contain

DANS at all,Again, just as the regulat 0888 991 orchw John io leavs~~produced thef;;ansiatlon "J'ordonne a Jean de partiP by using ghest%reotyPe the key "to", the less regular ease "I urge John tb

I cave” which requires the quite different oonstruttfon "J'exho;teJean a pattir" would be dealt with by a haltlno sie;eotyoe fo;ff u r g 0 " hhoee form :ouid be( EXHORTER (FNl MAN FOLK) (FNXl OBJE *DO) 9 A (FNXINF 430,)and in this case@ the stereotype for “to’ would n@v@r b@ cOnSUk@d atall'

FinallY ,It should be admltted that ln the aotual comPutaifbn of theanalYsk3 and generatIon system described above, two items of *informatlon I have desorlbedrcase and mark,shrink fnimportance,though bY no means dlsaPosar,~heit role has been

42

overstressed In the pa~ep,ln ordsr to make a oleai dfstlnoklonbetbeen the analysfs and Qsneratton roUt?ne8 and 80 pr@$eni a olarelntrrllngual rapresentatlon format,oPon to Inspeotion by any lln ujrtiunfafllllar wlth,and unlnterssted ln,the algbrlthmlo 3trchn ~~8semploYed,What I sought to avo1d wa8 any rallrtencre tb a w88am188scomputatlonal wholer~ all of whose i‘ov8ls seam to prosuppo80 all bithe other leWls,and whloh even If It works ,oannot be fn any wayfnspeotrd or d1soussed,

I hinted In the body of the paper that the ass~Qnm@nts bf ihr oassand mark lnformatton Itself demands aooess tb <ho FrsnohytereotWes,and li would ojsarly be absurd t6 oonsuli th8 sisrsoYYdssto assign this lnformatlon and then ,laterroon8ult ihsm aga n4 inorder to make use oi It In the generatlon of Ftonoh, I n fao&iheanalysis and oeneratlbn routlnes fu$e at this point ,and.tt(g caba andmark are located during the generation of the Frenoh output,

The change In the format that this reau1prs 1s that ihr markpredlcat@ PRMARK ia not now slmPlY a ared1oatr that'oheokt whrhhe;the ALREADY ASSIGNED mark for the fragment In hand meeti ihs8peclflcatlon:lt IS a predloate that at the same time aotlvelY seeksfor a mark meeting that soea1tlgatlon,And m with the siweotypsfunttlons a/ready dsscrlbed,the failure to find suoh a mark hfI8 ihswholo stereotype contalnlno It,there will now be not a slnglo markpred/cate, but a number of them fultllllno d/fterent rolrs,tho oa8epredloate,conVersely,fs not dlverslfled but vsstlg~al,b8oaure thsreIs now no PREVIOUSLY ASSIGNED aa8e ta a fraQmc, taheck,and the cau Is now Just a label c

for the pjedloate &n ihe dlotionary oi

stereotypes to ald ihe reader,

A aulak last IoOk ai a orevious s%ampIs rhould make all t h i selear,Conslder again (He hlt the boy Hwlth ihe wooden l8Q) a scontrasted with the qternatlve rreaond fragments (with a sjiok, and(wtth long halr),l.et ue oon8ldsr the rnalysls routines iermln?tlnoHfth t h e provision ot tull templates tar tra@msntr (and pha8elnformatton) ,and let us oonsldar everythlng that follows that a sFrenah generatlan,

Let- us now consider the Qenrratlon moOram enterfng the 88obndfta@ment,armed with the tollow1no lbst at 8fereotY~es fbr nwhhV

((PRMKOB *ENT)fPOSS) A (FN @ENT))((PRMARK ~DOHINST) AvEC (FN THING))((PRMARK +ENT)(POSS) A (FN *REAL),)

PRMKOB Is a dlreated predleateraa It wore,ihat 8@&8 for a mark fn apreceding fragment (w1thln a range of two fragments),It la&8 bnlv aioandldates whose heads are tn the class *ENT&at IS io layTMINGNMAN,FOLK,BEAST or WORLDtentlties fn some 8en8sparts+ the s a m e

that., oao have$wvbe the hsadr, 'ACT~TATEIPOINT e$c,ars n o t

attached to ward senses that ~8 oan Speak of h8 having party, PRMKQBoompares the formulas for potentlo mark8 fn the fhlrd&J8st,

43

posltlon of preceding fragments with the formula for the obJect InB the wtm late for the fragment In hand,And It Is true 1f and only If

c- the latter fomula jndtcates that It tlea to a word senss ihat oan bea part df the entity tied to the ‘Wndldate marW formula,

-

L

L

>

c-

So,ln the ease of (crs hit the bojO(wtth the wooden leg) PRMKOB finds;tse;f aomparfng the formulas for "bay" (head MANI and VW’ (whiahContaIna the sub-formula (MAN PART), In this asss PRMKOB ia saikffsdand the Oeneratton aontjnues through the fkt stereotYPe Oorrsc~lYgeneratlng ttatt for "w I th" and then the output for "woaden IsQ”,Ths*REAL In the function tn the tlrst stsreotyps msrsly lndfoates thaiany obJect In that fragment should then haha Its stereotype generaied(an9 substantive head I S in the class *REAL) baoaus@ 1 tsappropriatWwss has already been estabtlshed by the satisfaction ofPGMK08,~ol(awlng exactly the proaedurss dssorlbsd in other sW&ss Jk willbe seen that (with a stick) tails the first but 1s tranelatsd by <hesecond gtereotype,whl(e (with bW hair) tatis the ffrst two but ta

c- aorrsctly gene&ted by the third,

“b

c

L

b-

44

REFERENCES

l)Rar-Wll~l,Yc Some reflections on the present outlook for high-quality machine translation,MImeo,Unlv,of Texas, 1970,

2)Blerwisch,M, Semantics . .

in New Horizons In Llngulstlcs(ed,Lyons,J,) London.1970,

3)Kleln,S,et al, The Autolfng System,Tech,Report,#43,ComP,Sc/,peptlUn~v,of Wjsconsln,MadlsonrWlsc,1968,

4)Lakoff&, Llngutstlcs and Natural LoQ~,Studtes in CeneratfVe Semantics #i,Unlv,ofMlch\gan,Ann Arbor,Mloh,l970,

5)McCarthy,J,% Hayes ,P,Some PhIlosophIcal Problems from the strndpalntof Artificial Intelligence,in Machine 1ntel)lgenoa 4,Edlnburgh,l969

6)Michle,D, On not seeing things,--' ExperImental Ptogrammlng Reports #22

Un;'v,ot Edlnburgh,l971,7)Montague,R, Englfsh as a formal IanQuaQe

in Ltnguaggl nella Soclata e nella Teonlca,Milan 1970,

8)NldarE,6 Taberrc,The theory and Practloa of translation,Lefden 1969,

9)Qulllian,R, The teachable language comprehender,Comm,A,C& 1969,

10)Sandewall,E, Representlng natural language informationin predicate calculus,Machtne Intslllaence 6, Edinburgh 1971,

ll)Sohank,R, fhdfng the oon0ePtual conteni and lntrntlonof an utterance In natural lanfwew conversation,Proc,2nd,Jolnt Internattonal Conferenoe on A,!,London,l971,

~2WmmonsrR, Some SemantICI st+t;tures for +presentIng Englishmeanings,Tech Report #NL-l,Unlv,of Texae at Austln,1970,

13jkllksrYc On&e semantic analYsls of English texts,Mach,Trans,B Coma,Llng,l?68,

14)&llks,Y, Dec!dablilty and natural language,Mind, 1971"

15w Iks,Y, Grammar,meanfnQ and the maohlne analysis ofnatural language,London, 1971,

lb)klnograd,T, Procedures as a representation for data in a computerprogram for understanding natural language,ProJeot MAC Memo # MAC TRW84Massachusetts Institute of Teohnology11971'"

.